Power transmission channel foreign matter hidden danger high-precision rapid discrimination and alarm method

Abstract

The invention relates to the technical field of power transmission channel foreign matter identification, and discloses a power transmission channel foreign matter hidden danger high-precision rapid discrimination and alarm method, which comprises the steps of performing irregular foreign matter target feature extraction on video stream image data, performing feature matching based on an extracted irregular foreign matter target feature map, and executing foreign matter detection. And foreign matter state change trend perception is carried out on the irregular foreign matter detection result, foreign matter approximate distance estimation is carried out, grading alarm is carried out, and a foreign matter grading alarm result is obtained. The method solves the problem that a traditional algorithm is poor in detection rate of long-distance small targets and irregular foreign matters, improves the effectiveness of foreign matter alarm, effectively reduces the risk of external damage through precise recognition of floating foreign matters at the initial stage of hidden dangers, and improves the operation safety of a power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission channel foreign matter identification, in particular to a kind of power transmission channel foreign matter hidden danger high-precision rapid discrimination and warning method. BACKGROUND

[0002] Power transmission line dense channel has the characteristics of wide coverage, complex terrain, and severe natural environment, and the safety management and control of power transmission channel faces serious challenges. In recent years, due to the influence of factors such as intensive municipal construction, increased planting activities along the line, and frequent gale weather in the coastal area, plastic film, dust screen, color steel tile and other easily floating objects have caused continuous increase in power transmission line trip-out failures. The early stage of this power transmission channel has installed visual monitoring and video devices and other line online monitoring equipment, but the effect is limited and cannot meet the actual needs of quickly and accurately discovering foreign matter. Therefore, it is necessary to continue to find practical solutions.

[0003] As the "engine" technology in current artificial intelligence technology, deep learning technology has achieved results close to or even surpassing human level in image processing problems in many fields. Currently, the company has carried out some research and application in power grid inspection image processing technology, but has not considered the complexity of multiple types of foreign matter shapes, the complexity of inspection images, and the randomness of target relative motion. Therefore, there is still a lot of room for improvement in recognition accuracy, adaptability and generalizability. In the specific detection of foreign matter in power transmission channel, the core problems mainly include the following aspects:

[0004] 1. Extreme distribution of data samples and irregular shape of targets: due to the uncertainty of shape deformation, the collection of data for dust screen, tarpaulin and free foreign matter is difficult, and it is difficult to ensure the global applicability of existing samples to this type of target, which poses a challenge to model training. In addition, since traditional visual models mainly obtain feature information of a certain type of target object through training data set, non-rigid foreign matter has the characteristic of not obvious shape feature, which makes it difficult for the model to accurately capture the constantly changing shape of this type of target.

[0005] 2. High-precision identification of complex background and irregular foreign matter: the influence of background texture factors is significant, especially the overhead power transmission line inspection image often has background interference with similar characteristics to the target equipment. In addition, the target to be identified in the power transmission channel has a significant multi-scale three-dimensional structure feature. In the same image, there may be large targets with a size of several meters, or small targets with a size of only a few tens of centimeters. Limited by imaging distance and angle, there is occlusion between different targets or towers. Therefore, the high-precision identification algorithm of irregular foreign matter in complex background is also one of the key points and difficulties of the project research.

[0006] 3. Lack of self-iterative updating mechanism model: there is no closed loop of data collection, sample preprocessing, sample labeling, model training and model updating in the current power transmission channel scene, which leads to the inability of the algorithm to update itself regularly through the increasing training data.

[0007] However, no effective solution has been proposed to solve the above problems. SUMMARY

[0008] To overcome the above technical problems existing in the prior art, the present application proposes a high-precision and rapid identification and alarm method for foreign matter hidden dangers in a power transmission channel.

[0009] To this end, the specific technical solutions adopted by the present application are as follows:

[0010] A high-precision and rapid identification and alarm method for foreign matter hidden dangers in a power transmission channel includes:

[0011] Obtain video stream image data in the power transmission channel scene, and perform irregular foreign object feature extraction on the video stream image data, perform feature matching based on the extracted irregular foreign object feature map, perform foreign object detection according to the feature matching result, and obtain the irregular foreign object detection result;

[0012] Perform foreign object state change trend perception on the irregular foreign object detection result, and perform foreign object approximate distance estimation according to the foreign object state change trend perception result, perform hierarchical alarm based on the foreign object approximate distance estimation result, and obtain the foreign object hierarchical alarm result.

[0013] Further, the irregular foreign object feature extraction on the video stream image data, the feature matching based on the extracted irregular foreign object feature map, and the irregular foreign object detection result obtained according to the feature matching result include:

[0014] Perform augmentation and enhancement processing on the video stream image data to obtain enhanced video stream image data, and perform irregular foreign object feature extraction on the enhanced video stream image data based on the irregular convolution kernel of deformable convolution, and obtain the irregular foreign object feature map;

[0015] Perform feature matching on the irregular foreign object feature map using a hybrid domain attention mechanism to obtain a feature matching result;

[0016] Perform feature extraction based on semantic segmentation on the feature matching result based on a hollow spatial convolution to obtain a dense feature representation, and perform feature enhancement based on an attention mechanism on the dense feature representation to obtain an irregular foreign object detection result.

[0017] Further, the video stream image data is subjected to augmentation and enhancement processing to obtain enhanced video stream image data, and the irregular foreign object target feature map is extracted from the enhanced video stream image data based on an irregular convolution kernel of deformable convolution, and the irregular foreign object target feature map includes:

[0018] The video stream image data is subjected to data augmentation processing to obtain augmented video stream image data, and the augmented video stream image data is subjected to image aliasing-based data enhancement processing to obtain enhanced video stream image data;

[0019] The enhanced video stream image data is input into a deep convolutional neural network of deformable convolution, and preliminary feature extraction is performed using an irregular convolution kernel to obtain an irregular foreign object preliminary feature map;

[0020] The offset amount in the coordinate direction of each sampling point of the irregular foreign object preliminary feature map is calculated using a bias layer, and a bias matrix is generated according to the offset calculation result;

[0021] Based on the sampling point offset amount in the bias matrix, the irregular foreign object preliminary feature map is updated and optimized using an interpolation algorithm and a back propagation algorithm to obtain an irregular foreign object target feature map.

[0022] Further, the irregular foreign object target feature map is subjected to feature matching using a mixed domain attention mechanism to obtain a feature matching result, which includes:

[0023] The irregular foreign object target feature map is input into a channel domain attention mechanism, and parallel maximum pooling and average pooling operations are performed based on the width and height of the irregular foreign object target feature map to obtain two extruded foreign object target feature maps;

[0024] The two extruded foreign object target feature maps are subjected to addition operation using a multi-layer perceptron, and a normalized channel attention feature is generated through a sigmoid activation function;

[0025] The channel attention feature and the irregular foreign object target feature map are subjected to multiplication operation to obtain an input feature map of a spatial domain attention mechanism;

[0026] Global maximum pooling and average pooling serial operations are performed on the input feature map to obtain two serial operation feature maps, the two serial operation feature maps are spliced, and a spatial dimension attention feature is generated using convolution and sigmoid activation function;

[0027] The spatial dimension attention feature is weighted to the irregular foreign object target feature map through multiplication operation to obtain the feature matching result.

[0028] Further, based on the cavity space convolution, the feature matching result is subjected to semantic segmentation-based feature extraction to obtain a dense feature representation, and the dense feature representation is subjected to attention mechanism-based feature enhancement to obtain an irregular foreign matter detection result including:

[0029] The feature matching result is subjected to three downsampling operations to obtain a low-dimensional feature map, and the channel number of the low-dimensional feature map is halved and divided into two parallel branches;

[0030] Each parallel branch uses a depth separable convolution instead of a standard convolution, and a cavity convolution with different cavity rates is used in the convolution layer of each parallel branch to obtain a dense feature representation that fuses multi-scale context information;

[0031] The dense feature representation is subjected to channel and spatial information expression enhancement through channel attention and spatial attention to obtain attention branch features, and a global average pooling branch is introduced to integrate global context information to obtain global context branch features;

[0032] The attention branch features and the global context branch features are fused, and the resolution of the feature matching result is matched through an upsampling operation to obtain an irregular foreign matter detection result.

[0033] Further, the irregular foreign matter detection result is subjected to foreign matter state change trend perception, and foreign matter approximate distance estimation is performed according to the foreign matter state change trend perception result, based on the foreign matter approximate distance estimation result, hierarchical warning is implemented to obtain foreign matter hierarchical warning result including:

[0034] Based on the TransT tracking algorithm, the feature fusion processing is performed on the irregular foreign matter detection result to obtain a fusion feature response map, and the foreign object tracking positioning processing is performed on the fusion feature response map to obtain an irregular foreign matter motion trajectory;

[0035] The change trend predictor is used to perform foreign matter state change trend perception based on time series data on the irregular foreign matter motion trajectory to obtain a foreign matter state change trend perception result;

[0036] Based on the foreign matter state change trend perception result, the foreign matter approximate distance estimation is performed based on semantic segmentation to obtain a foreign matter approximate distance estimation result;

[0037] The foreign matter approximate distance estimation result is combined with the preset warning threshold rule to perform foreign matter hierarchical warning to obtain a foreign matter hierarchical warning result.

[0038] Further, based on the TransT tracking algorithm, the feature fusion processing is performed on the irregular foreign matter detection result to obtain a fusion feature response map, and the foreign object tracking positioning processing is performed on the fusion feature response map to obtain an irregular foreign matter motion trajectory including:

[0039] The TransT tracking algorithm based on a twin network structure is used to extract features from the non-regular foreign object detection results, to obtain template features and search area features.

[0040] The template features and search area features are input into a feature fusion network, and the integration of different position information is implemented based on context enhancement and cross feature enhancement of the multi-head attention mechanism, to obtain an integration result.

[0041] The pixel point space position of the integration result is encoded using a sine function to obtain a fusion feature response map.

[0042] The fusion feature response map is input into a classification branch and a regression branch respectively, and the classification output and the regression output of the current frame are generated, and the highest anchor frame of the foreground classification score is filtered by the tracker to obtain the non-regular foreign object motion trajectory.

[0043] Further, the change trend predictor is used to perform foreign object state change trend perception based on time series data on the non-regular foreign object motion trajectory, to obtain a foreign object state change trend perception result including:

[0044] When the non-regular foreign object motion trajectory is in the first frame, the change trend predictor based on Kalman filtering is initialized by the tracker to obtain an initial state vector.

[0045] The state prediction equation is used to make a priori estimate of the state variable to obtain a priori estimate value, and the priori estimate value of the change trend predictor is corrected based on the tracking frame of the tracker, and the change trend predictor parameters are iteratively updated to obtain an updated change trend predictor.

[0046] The non-regular foreign object motion trajectory is predicted and updated based on the updated change trend predictor to obtain a foreign object state change trend perception result.

[0047] Further, the foreign object approximate distance estimation based on semantic segmentation is implemented according to the foreign object state change trend perception result to obtain a foreign object approximate distance estimation result including:

[0048] The identified object of semantic segmentation is determined according to the foreign object state change trend perception result, and pixel-level classification recognition is implemented based on the identified object to obtain a pixel-level classification recognition result.

[0049] The spatial relationship between the two is judged based on the pixel-level classification recognition result, and the spatial relationship judgment result is normalized to obtain a foreign object approximate distance estimation result.

[0050] Further, the grading determination rule of the grading alarm is: when the foreign matter approximate distance estimation result is greater than or equal to 0.3, it is classified as a first-level alarm; when the foreign matter approximate distance estimation result is less than or equal to 0.1, it is classified as a second-level alarm; when the foreign matter approximate distance estimation result is greater than 0.1 and less than 0.3, then the alarm level is determined again in combination with the foreign matter type and the weather condition.

[0051] The beneficial effects of the present application are:

[0052] 1. The present application improves the adaptability of the deep neural network to the complex power transmission environment through the high-precision foreign matter detection algorithm, solves the problem of poor detection rate of traditional algorithms for long-distance small targets and irregular foreign matters, improves the effectiveness of foreign matter alarm, and helps the stable, safe and efficient operation of the power transmission line.

[0053] 2. The present application highly accurately locates the safety hazard points and fault points by intelligently analyzing the power transmission channel images, and then guides the maintenance team to quickly carry out line maintenance and abnormal processing, which can reduce the labor intensity of the maintenance workers, shorten the inspection cycle, improve the power transmission line and substation operation and maintenance capacity of the company, provide strong information technology support for the stable operation of the power grid, and comprehensively improve the intelligentization and lean management of the power grid business; further makes up for the shortcomings of traditional artificial intelligence technology in the use of power transmission and transformation equipment inspection process, and marks the landing and implementation of the new generation of artificial intelligence technology in the power field.

[0054] 3. The present application focuses on the high-precision real-time identification and early warning of floating foreign matters in the hazard stage in the power transmission channel, breaks through the passive response mode of traditional anti-external damage alarm technology which is only aimed at foreign matters that have been suspended or after external damage occurs, realizes the transformation from "after-treatment" to "prevention", and proposes a significant differentiated technical positioning from traditional anti-external damage alarm, that is, by accurately identifying floating foreign matters in the early stage of hazards, such as kites, plastic films, etc., the alarm is triggered before the floating foreign matters contact the conductor, effectively reducing the external damage risk and improving the safety of power grid operation. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0056] Figure 1 is a flowchart of the high-precision and rapid identification and alarm method for foreign matter hazards in the power transmission channel according to the embodiments of the present application;

[0057] Figure 2It is a deformable convolution schematic diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0058] Figure 3 It is an implementation process diagram of deformable convolution in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0059] Figure 4 It is a mixed attention mechanism schematic diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0060] Figure 5 It is a segmentation network overall framework diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0061] Figure 6 It is a TransT algorithm structure schematic diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0062] Figure 7 It is a Kalman filter-based change trend predictor cycle process schematic diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application;

[0063] Figure 8 It is a multi-source information fusion alarm strategy diagram in the power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application. DETAILED DESCRIPTION

[0064] To further illustrate the embodiments, the present application provides drawings, which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can explain the operating principle of the embodiments in conjunction with the related description of the specification. With reference to these contents, those skilled in the art should understand other possible embodiments and advantages of the present application. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0065] According to the embodiment of the present application, a power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method is provided.

[0066] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figures 1-8 The power transmission channel foreign matter hidden danger high-precision rapid identification and alarm method according to the embodiment of the present application comprises:

[0067] The video stream image data in the power transmission channel scene is acquired, and non-regular foreign object target feature extraction is performed on the video stream image data. Feature matching is performed based on the extracted non-regular foreign object target feature map, and foreign object detection is performed according to the feature matching result to obtain a non-regular foreign object detection result.

[0068] Specifically, the non-regular foreign object target feature extraction is performed on the video stream image data, the feature matching is performed based on the extracted non-regular foreign object target feature map, and the foreign object detection is performed according to the feature matching result to obtain the non-regular foreign object detection result, which includes:

[0069] The video stream image data is subjected to augmentation and enhancement processing to obtain enhanced video stream image data, and non-regular foreign object target feature extraction is performed on the enhanced video stream image data based on a non-regular convolution kernel of deformable convolution to obtain a non-regular foreign object target feature map.

[0070] Specifically, the non-regular foreign object target feature extraction is performed on the video stream image data, the feature matching is performed based on the extracted non-regular foreign object target feature map, and the foreign object detection is performed according to the feature matching result to obtain the non-regular foreign object detection result, which includes:

[0071] The video stream image data is subjected to data augmentation processing to obtain augmented video stream image data, and the augmented video stream image data is subjected to data enhancement processing based on image aliasing to obtain enhanced video stream image data.

[0072] The enhanced video stream image data is input into a deep convolutional neural network of deformable convolution, and a non-regular foreign object preliminary feature map is obtained by using a non-regular convolution kernel to perform preliminary feature extraction.

[0073] The offset amount of each sampling point in the coordinate direction of the non-regular foreign object preliminary feature map is calculated by using a bias layer, and a bias matrix is generated according to the offset amount calculation result.

[0074] Based on the sampling point offset amount in the bias matrix, the non-regular foreign object preliminary feature map is updated and optimized by using an interpolation algorithm and a back propagation algorithm to obtain a non-regular foreign object target feature map.

[0075] The feature matching of the non-regular foreign object target feature map is performed by using a mixed domain attention mechanism to obtain a feature matching result.

[0076] Specifically, the feature matching of the non-regular foreign object target feature map is performed by using a mixed domain attention mechanism to obtain a feature matching result, which includes:

[0077] The non-regular foreign object target feature map is input into a channel domain attention mechanism, and parallel operations of maximum pooling and average pooling are performed based on the width and height of the non-regular foreign object target feature map to obtain two squeezed foreign object target feature maps.

[0078] The two extruded foreign object feature maps are added by using a multi-layer perception, and normalized channel attention features are generated by using a sigmoid activation function;

[0079] The channel attention features are multiplied with the irregular foreign object feature map to obtain an input feature map of the spatial domain attention mechanism;

[0080] Serial operations of global max pooling and average pooling are performed on the input feature map to obtain two serial operation feature maps, the two serial operation feature maps are spliced, and a spatial dimension attention feature is generated by using convolution and a sigmoid activation function;

[0081] The spatial dimension attention feature is weighted channel by channel on the irregular foreign object feature map by multiplication to obtain a feature matching result.

[0082] Based on the cavity space convolution, semantic segmentation-based feature extraction is performed on the feature matching result to obtain a dense feature representation, and attention mechanism-based feature enhancement is performed on the dense feature representation to obtain an irregular foreign object detection result.

[0083] Specifically, based on the cavity space convolution, semantic segmentation-based feature extraction is performed on the feature matching result to obtain a dense feature representation, and attention mechanism-based feature enhancement is performed on the dense feature representation to obtain an irregular foreign object detection result, which includes:

[0084] Three downsampling operations are performed on the feature matching result to obtain a low-dimensional feature map, the number of channels of the low-dimensional feature map is halved and divided into two parallel branches;

[0085] Each parallel branch uses a depth separable convolution instead of a standard convolution, and a cavity convolution with different cavity rates is used in the convolution layer of each parallel branch to obtain a dense feature representation that fuses multi-scale context information;

[0086] The dense feature representation is enhanced in terms of channel and spatial information expression by channel attention and spatial attention to obtain an attention branch feature, and a global average pooling branch is introduced to integrate global context information to obtain a global context branch feature;

[0087] The attention branch feature and the global context branch feature are fused, and the resolution of the feature matching result is matched by upsampling operation to obtain an irregular foreign object detection result.

[0088] The irregular foreign object detection result is subjected to foreign object state change trend perception, and foreign object approximate distance estimation is performed according to the foreign object state change trend perception result, and hierarchical warning is performed based on the foreign object approximate distance estimation result to obtain a foreign object hierarchical warning result.

[0089] Specifically, the irregular foreign matter detection result is subjected to foreign matter state change trend perception, and foreign matter approximate distance estimation is performed according to the foreign matter state change trend perception result, hierarchical warning is implemented based on the foreign matter approximate distance estimation result, and the foreign matter hierarchical warning result includes:

[0090] The irregular foreign matter detection result is subjected to feature fusion processing based on the TransT tracking algorithm, a fusion feature response graph is obtained, and foreign object tracking positioning processing is performed on the fusion feature response graph to obtain an irregular foreign matter motion trajectory.

[0091] Specifically, the irregular foreign matter detection result is subjected to feature fusion processing based on the TransT tracking algorithm, a fusion feature response graph is obtained, and foreign object tracking positioning processing is performed on the fusion feature response graph to obtain an irregular foreign matter motion trajectory, including:

[0092] The TransT tracking algorithm based on a twin network structure is used to extract features from the irregular foreign matter detection result to obtain template features and search region features.

[0093] The template features and the search region features are input into a feature fusion network, and the integration of different position information is implemented based on the context enhancement and cross-feature enhancement of the multi-head attention mechanism to obtain an integration result.

[0094] The pixel point space positions of the integration result are encoded using a sine function to obtain a fusion feature response graph.

[0095] The fusion feature response graph is input into a classification branch and a regression branch, respectively, and classification output and regression output of the current frame are generated, and the highest anchor frame of the foreground classification score is filtered through a tracker to obtain an irregular foreign matter motion trajectory.

[0096] The change trend predictor is used to perform foreign matter state change trend perception based on time series data on the irregular foreign matter motion trajectory to obtain a foreign matter state change trend perception result.

[0097] Specifically, the change trend predictor is used to perform foreign matter state change trend perception based on time series data on the irregular foreign matter motion trajectory to obtain a foreign matter state change trend perception result, including:

[0098] When the irregular foreign matter motion trajectory is in the first frame, the change trend predictor based on Kalman filtering is initialized using the tracker to obtain an initial state vector.

[0099] The state prediction equation is used to perform prior estimation on the state variable to obtain a prior estimate value, the prior estimate value of the change trend predictor is corrected based on the tracking frame of the tracker, and the change trend predictor parameters are iteratively updated to obtain an updated change trend predictor.

[0100] The non-regular foreign matter motion trajectory is predicted and updated based on the updated change trend predictor, and a foreign matter state change trend perception result is obtained.

[0101] The foreign matter approximate distance estimation based on semantic segmentation is implemented according to the foreign matter state change trend perception result, and a foreign matter approximate distance estimation result is obtained.

[0102] Specifically, the foreign matter approximate distance estimation based on semantic segmentation is implemented according to the foreign matter state change trend perception result, and a foreign matter approximate distance estimation result is obtained, including:

[0103] The identified object of semantic segmentation is determined according to the foreign matter state change trend perception result, and a pixel-level classification recognition result is obtained based on the identified object.

[0104] The spatial relationship between the two is judged based on the pixel-level classification recognition result, and the spatial relationship judgment result is normalized to obtain the foreign matter approximate distance estimation result.

[0105] The foreign matter approximate distance estimation result is combined with the preset alarm threshold rule to perform foreign matter grading alarm, and a foreign matter grading alarm result is obtained.

[0106] Specifically, the grading determination rule of the grading alarm is:

[0107] When the foreign matter approximate distance estimation result is greater than or equal to 0.3, it is classified as a first-level alarm.

[0108] When the foreign matter approximate distance estimation result is less than or equal to 0.1, it is classified as a second-level alarm.

[0109] When the foreign matter approximate distance estimation result is greater than 0.1 and less than 0.3, a secondary determination of the alarm level is implemented in combination with the foreign matter type and the weather condition.

[0110] In order to facilitate the understanding of the above technical solutions of the present application, the working principle or operation mode of the present application in the actual process will be described in detail.

[0111] In practical application, for the non-regular target feature extraction technology for high-precision detection of non-regular foreign matters in complex backgrounds, image expansion is mainly performed through conventional expansion and image aliasing.

[0112] 1. Conventional data set expansion: the target features extracted by the convolutional neural network are fixed, so the data set can be expanded using data enhancement. The common data augmentation methods in the field of power target detection include flipping, rotating, scaling, cropping, adding Gaussian noise, and brightness adjustment, etc. (1) Flipping: mainly including horizontal flipping and vertical flipping. Vertical flipping is equivalent to rotating the original picture by 180° around the center, and the target detection object is the same as the original, so horizontal flipping is mainly used to enhance the foreign object picture. Original image and image after flipping. (2) Rotation: for rectangular input pictures, in addition to the rotation operation around the center angle of 180°, the rest will change the dimensions of the picture. A rotation operation around the center by 90° is used for data enhancement. (3) Scaling: scaling operation is divided into outward and inward. After outward, the size of the picture will become larger, and it needs to be cut to the original size after transformation. The effect of small-scale outward operation on the image, since the outward scale is small, the difference between the original image is not large in vision. (4) Cropping: cropping is to randomly crop a region in the original image, and then adjust it to the size of the original picture. Cropping operation can enlarge the smaller detection target in the original image, simulate the detection of the same object under different shooting distances, and also simulate the incomplete appearance of the object to be detected due to occlusion, thereby enhancing the learning ability and target discrimination ability of the model. (5) Gaussian noise: adding Gaussian noise in training can simulate the situation that the image is disturbed by noise. Gaussian noise can also be used to simulate weather environments such as sand and dust, and improve the learning ability of the model in harsh environments. The effect of adding Gaussian noise. (6) Brightness adjustment: since the power line is exposed outdoors, the lighting conditions are uncertain, so brightness adjustment can be used to simulate different lighting environments. By adjusting the brightness, the high-intensity light environment can be simulated, and vice versa to simulate low-intensity light environment to enhance the learning ability of the model. Image after brightness adjustment processing.

[0113] 2. Image Mosaic-based data augmentation: In order to better obtain feature information, the size of the input image needs to be changed, but this operation will increase the computational load of the model and consume more memory and resources. If the downsampling operation is limited, it is not conducive to extracting deep information. In order to solve this problem, image mosaic-based data augmentation methods such as Mosaic and Mixup are proposed to further expand the data set, which can improve the recognition rate with only a small amount of additional computation. (1) Mosaic data augmentation: Mosaic refers to randomly selecting four images in the data set, scaling them, and then arranging them in a left-up, left-down, right-up, and right-down manner to splice them together to get a new image with uniform size. The new image contains more background information than the original image, and the splicing operation also increases the number of samples (batch size) passed to the model for training at a time, which reduces the dependence on the batch size itself when calculating the batch normalization (Batch Normalization).

[0114] Convolutional neural networks rely on convolutional layers to extract features from input images. Traditional convolution generally uses a fixed size and proportion of rectangular structures to extract features at a specific location of the feature map. In order to make the convolutional neural network have better ability to recognize geometric transformations, deformable convolution DCN (Deformable Convolution) is introduced instead of regular convolution. Deformable convolution can automatically adapt and adjust according to the shape and proportion of the object to be measured, and by using irregular convolution kernels, it solves the problem of insufficient sampling of traditional fixed rectangular structures, so that the network model can better simulate the deformation of objects. The principle of deformable convolution is shown in Figure 2 The principle of deformable convolution is to add an offset to each sampling point based on traditional convolution. Assuming the size of the convolution kernel is 3x3, there are 9 sampling points for this convolution kernel, and researchers assign an offset to each of the 9 sampling points. Therefore, the position and size of the convolution kernel can be dynamically adjusted according to the target object, which can better fit the size and shape of the object during feature extraction, optimizing the detection performance of the network model for irregular and severely deformed objects. For traditional two-dimensional standard convolution, the formula for the output feature value of a certain sampling point p0 is as follows: In the above formula, R = {(-1, -1), (-1, 0), (0, 1), (1, 1)}, w(p n ) is the convolution kernel weight value at the position of p n , x is the feature map to be detected, and p0 represents the position element in R, x(p n+ p0) represents the sampling position of the feature map to be tested. In deformable convolution, the calculation formula of the output feature value y(p0) is as follows: In the above formula, Δp n is the offset added on the regular convolution sampling point. As Figure 3 shown, first, the same as the traditional convolutional neural network, the input image is subjected to feature extraction by using the regular convolution kernel. Then the output feature map obtained is subjected to an additional offset field layer to obtain the offset field of the convolution kernel sampling point, and the number of channels thereof is 2N, because the number of channels includes the offsets in the x and y directions on the two-dimensional plane. The size of the offset field obtained and the input feature Figure 1 map can be used to obtain the offset matrix of the convolution kernel sampling point, so that the offset Δp n can be obtained. The sampling point offset is updated and optimized by using the interpolation algorithm back propagation algorithm, and is generally a decimal number. The introduction of deformable convolution into the network model not only improves the fitting ability of irregular shaped objects, but also reduces the omission during feature extraction, so that the target information of the measured object is more complete. Figure 3 The implementation process of deformable convolution is shown. The input feature map is first subjected to a standard convolution layer (conv) to extract features and generate an offset field, the number of channels of which is 2N, for representing the offsets in the x and y directions of each sampling point. These offsets are superimposed on the fixed sampling positions of the traditional convolution kernel to form dynamically adjusted sampling points, thereby realizing adaptive sampling of irregular or deformed targets in the input feature map. Subsequently, after the deformable convolution operation, the output feature map can more accurately capture the geometric changes and spatial structure information of the target. This mechanism effectively improves the expression ability of the model for irregular shaped objects, and is particularly suitable for target detection tasks of objects with severe deformation such as floating objects in the transmission channel.

[0115] For the complex background foreign matter high-precision feature matching algorithm in the complex background foreign matter high-precision detection, a hybrid domain attention mechanism is adopted to realize the complex background foreign matter high-precision feature matching. The hybrid domain attention mechanism mainly includes a channel domain attention mechanism and a spatial domain attention mechanism. The channel domain attention refers to giving different weights to the features of different channels in the channel direction through the operation in the neural network model training, to represent the attention degree of the channel region. The spatial domain attention refers to giving different weights to different positions of the feature map through the operation in the neural network model training, to represent the attention degree of the position region. The hybrid domain attention mechanism is specifically as followsFigure 4 The mechanism acts on the channel domain and the spatial domain of the image at the same time, adjusts the direction of weight optimization in the two dimensions, and makes the network focus on the extraction of more important information. CBAM designs two independent sub-modules, which are the channel attention (CA) module and the spatial attention (SA) module. When the feature map is input, it first enters the CA module, and parallel operations of maximum pooling and average pooling based on the width and height of the feature map are performed to realize the squeezing of the image dimension. Compared with a single pooling operation, parallel pooling can obtain two 1x1xC feature maps, which retain more information. After the two feature maps pass through the shared multi-layer perception, they are added and operated, and the normalized channel attention feature is obtained through the sigmiod activation function. The feature is multiplied with the input feature map to generate the input feature required by the SA module. In the SA module, first, the input feature is subjected to serial operations of global maximum pooling and average pooling based on the channel to obtain two HxWx1 feature maps. The two feature maps are spliced, and then a 7x7 convolution and a sigmoid activation function are used to obtain the attention feature of the spatial dimension. The attention feature is weighted to the original input feature map through multiplication operation, i.e. the feature map is redefined, to obtain the final output feature map, which can effectively avoid the influence of the background on the foreign object feature and realize high-precision matching of the foreign object feature.

[0116] For the foreign object detection technology based on video stream semantic segmentation in high-precision detection of irregular foreign objects in complex background, to balance the high algorithm precision and fast operation speed, the irregular target feature extraction technology and the complex background foreign object high-precision feature matching algorithm are combined to obtain different size receptive field information. A combination of depth separable convolution and different size and rate of dilated convolution module is designed. The depth separable convolution and the dilated convolution with different rates of dilated convolution are combined to reduce the model calculation amount while more efficiently extracting the features. The channel attention module is introduced to obtain the importance between different channels of the features. Then, the spatial attention module is combined in a cascaded manner to enhance the expression of the spatial information and the channel information of the features, and the original features are fused to further improve the expression ability of the features. A real-time semantic segmentation network based on dilated spatial convolution pyramid and attention mechanism is proposed. The schematic diagram of the real-time semantic segmentation network based on dilated spatial convolution pyramid and attention mechanism is shown in Figure 5The feature extraction stage is composed of a down-sampling module and a dilated separable convolution module. The feature enhancement stage is composed of a channel attention module, a spatial attention module, a global average pooling, and an up-sampling. In the feature extraction stage, the input image is first subjected to a down-sampling operation, which is composed of a 3x3 convolution with a step of 2 and a 2x2 max-pooling. Because the down-sampling operation reduces the size of the feature map, it has an impact on the final prediction result. In order to preserve the spatial information details, the model first uses 3 times of down-sampling operation to obtain the feature map resolution of the original Figure 1 / 8, and then uses the designed dilated separable convolution module to densely extract features. This module uses a channel separation operation to divide the channel number into half of the original, and divides it into two branches. Each branch uses a depth separable convolution instead of a standard convolution to reduce the number of parameters, and uses a dilated convolution with different dilated rates in the convolution layer of each branch to expand the receptive field and effectively obtain multi-scale context information. In the feature enhancement stage, the extracted features are re-integrated and the feature information expression is enhanced. The specific operation is as follows: first, the channel attention module and the spatial attention module branch are introduced into the model at the same time to enhance the expression of channel information and spatial information of the features. Secondly, in order to further improve the performance of semantic segmentation, a global average pooling branch is introduced to integrate the global context information. Finally, all branch features are fused and an up-sampling operation is used to match the resolution of the input image. When down-sampling, part of the channels are appropriately cropped to reduce the amount of calculation and improve the inference speed of the model, and the accuracy does not decrease significantly.

[0117] Time and space-based foreign object hazard warning mechanism based on time series data-based foreign object state change trend perception technology:

[0118] 1) Irregular Object Tracking and Recognition Algorithm Based on Time-Series Data: This tracking and recognition algorithm operates on video or image sequences. Its main task is to process the target given in the first frame of the image sequence, extract effective features, and use the extracted features and contextual information in the video or image sequence to calculate the target's exact position in the next frame, ultimately obtaining the target's complete trajectory. The tracker needs to continuously update the target's features, acquiring features that are as close as possible to the target's current state and as accurate as possible to ensure the tracker can adapt to changes in the target. The tracker also needs to reasonably adjust the weights of the target's color features, structural features, boundary features, and other feature types to reduce the impact of illumination on the tracker's accuracy. Since the beginning of target tracking research, based on the different modeling methods used by the tracker, it can be mainly divided into generative target tracking methods and discriminative target tracking methods. Generative methods have many limitations compared to discriminative methods, while discriminative methods are currently the mainstream method for target tracking. These methods are mainly divided into three types: correlation filtering methods, Siamese network-based methods, and Transformer-based methods. A novel attention-based feature fusion network, TransT, is proposed using a Transformer-based target tracking algorithm. This network is integrated with a Siamese network structure, such as... Figure 5 As shown. The TransT algorithm still uses the two shared parameter branches of the traditional Siamese network in the feature extraction part. Then, TransT replaces the cross-correlation operation of the traditional Siamese network with a newly proposed feature fusion network. After fusing features, the tracker can obtain a fused feature response map. The fused feature response map is input into the classification branch and the regression branch respectively to obtain the classification output and regression output of the current frame. Finally, the tracker will take the anchor box with the highest foreground classification score as the final output. The most prominent contribution of the TransT algorithm is its proposed feature fusion network, which adopts a multi-head attention mechanism. The feature fusion network mainly consists of two modules, namely the context enhancement module (ECA) and the cross feature enhancement module (CFA). The context enhancement module adopts residual form multi-head self-attention, adaptively integrating information from different positions in the feature map. At the same time, since the attention mechanism is unable to distinguish the positional information of the input feature sequence, TransT also introduces a spatial position encoding to locate the recombined coordinates. The working mechanism of the context enhancement module can be summarized as: X EC =X+MultiHead(X+P x ,X+P x P(X); where X represents the input of the context enhancement module, P(X) ... x Represents spatial location encoding, X EC This is the output of the context enhancement module. The TranT method uses a sine function to encode the spatial location of pixels.

[0119] 2) Foreign matter change trend prediction algorithm based on time series data:

[0120] The change trend prediction problem, as an important part of computer vision research, also has a wide range of applications in power line scenes. The change trend prediction problem can be simply summarized as predicting the possible position of the target in the next frame through the historical coordinates and the current state of the target. Kalman filter is used as a change trend predictor, and the tracking results of the tracker are used to initialize and update the change trend predictor. First, the tracker is defined as a system, and at time t, the state variable of the tracker is x t =(cx t ,cy t ) T , that is, the center point coordinates (cx t ,cy t ) of the t-th frame tracker prediction box. For target tracking in any scene, the state variable x t and the observation variable Z t can be represented as:

[0121]

[0122] where w t-1 represents the process noise, which is a random noise value conforming to the multivariate normal distribution with mean 0 and covariance matrix Q, A t represents the state transition matrix, B t is the control matrix, u t is the control variable of the model, H is the state observation matrix, v t is the observation noise, and like the process noise, it is a random noise subject to the multivariate normal distribution with mean 0 and covariance matrix R. The process noise and the observation noise are independent of each other. The state variable in the formula is the system state predicted by the change trend predictor according to the historical experience and the current state, while the observation variable is the predicted value calculated by the tracker in the actual tracking process, that is, the "observation" obtained by the change trend predictor in the tracker system.

[0123] When the tracking process is in the first frame, that is, at time t=0, the tracker initializes the change trend predictor, and the center point coordinates of the target real value box given in the first frame are taken as the initial state vector of the change trend predictor, that is, x0=(cx0,cy0) TIn the second frame and subsequent frames of the video sequence, the change trend predictor will predict the system state, i.e. the target center position, while updating the change trend predictor parameters with the predicted box center coordinates of the tracker. For the t-th frame in the video sequence, the change trend predictor makes a priori estimate of the state variable through the state prediction equation. After the change trend predictor predicts the target position of the next frame, the predictor also needs to correct the prior estimate value of the change trend predictor using the tracking box obtained by the tracker, and update the related parameters of the change trend predictor. This process corrects the predicted estimate value of the change trend predictor by the observation value, and updates the parameters in the change trend predictor according to the observation value, so that it is as close as possible to the actual state of the system. The "prediction-update" cycle of the change trend predictor is shown in Figure 7 The core cycle of the Kalman filter algorithm is shown in FIG. 1; it shows the core cycle of the Kalman filter algorithm, which is divided into two stages of prediction and update. In the prediction stage, the prior state estimation is made according to the state estimation of the last time, and the corresponding prior error covariance matrix is calculated to predict the state at the current time; then enter the update stage, use the difference between the observation value and the predicted value, adjust the prediction result by calculating the Kalman gain, and then get more accurate posterior state estimation and posterior error covariance matrix; finally, the updated state is taken as the input of the next period to form a closed loop iteration, realizing continuous and robust tracking and estimation of the target state.

[0124] The intelligent monitoring alarm level classification strategy based on foreign object type, weather condition, distance, change trend and other multi-source data in the foreign object hidden danger alarm mechanism based on time and space factors:

[0125] 1) Foreign object approximate distance estimation method for power line channel scene:

[0126] Currently in the field of computer vision, the method closest to the approximate distance estimation method is the depth estimation method. The depth estimation method is to estimate the distance between the target object and the camera, while the approximate distance estimation method refers to the distance estimation between the target objects. In view of the scene of the power transmission line, most of the pre-installed monocular ordinary cameras, due to the irregular shape of the floating object in the power transmission channel, after imaging by the camera, the specific width and height information of the floating object is not known, in order to estimate the distance between the floating object and the conductor, the shape information of the conductor and the floating object is converted into a pixel area in the image. Based on the above, a monocular camera approximate distance estimation method based on semantic segmentation is proposed. In the monocular camera approximate distance estimation method based on semantic segmentation, the conductor and the floating object are both recognized as semantic segmentation objects, and the recognition result is a pixel-level classification. If the pixel area of the conductor and the pixel area of the floating object intersect, it is considered that the distance between them is 0.0. If there is only a conductor or only a floating object pixel area in the result image, it is considered that the distance is 1.0. In other cases, the distance between the floating object and the conductor is normalized to 0-1 according to the overall proportion of the image size.

[0127] 2) Intelligent monitoring alarm level classification strategy based on multi-source data:

[0128] The traditional monitoring system alarm signal mainly relies on a single information source, and cannot distinguish the level or urgency of the alarm signal. According to the type of foreign matter, weather conditions, distance, change trend and other conditions, an intelligent monitoring alarm level classification strategy based on multi-source data is proposed. The alarm strategy flow chart is shown in Figure 8 When the recognition algorithm detects a floating object, it is first tracked using a tracking algorithm and the change trend is predicted. When the change trend is away from the power transmission line conductor, this alarm is classified as a prompt alarm. When the change trend is approaching, the monocular camera approximate distance estimation method based on semantic segmentation is used to estimate the approximate distance of the floating object from the conductor. When the distance is greater than or equal to 0.3, it is classified as a general alarm. When the distance is less than or equal to 0.1, it is directly classified as a serious alarm. When the distance is greater than 0.1 and less than 0.3, it is determined whether the foreign matter is a color steel tile or a conductive material. If it is a conductive material, it is classified as a serious alarm. When it is an insulating material, the weather conditions are determined. When the weather is bad, it is classified as a serious alarm. When the weather is good, this alarm is classified as a general alarm.

[0129] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "screwing" and other terms should be understood in a broad sense, for example, can be fixedly connected, or can be detachably connected, or can be integrated; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium; can be internal communication of two elements or interaction relationship between two elements, unless otherwise explicitly limited, the above-mentioned terms in the present application can be understood according to the specific meaning of the above-mentioned terms in the present application by those skilled in the art.

[0130] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for high-precision and rapid identification and warning of foreign matter hazards in a power transmission channel, characterized in that, The method comprises the following steps: acquiring video stream image data in a power transmission channel scene, performing irregular foreign object target feature extraction on the video stream image data, performing feature matching based on the extracted irregular foreign object target feature map, performing foreign object detection according to the feature matching result, and obtaining an irregular foreign object detection result; performing foreign object state change trend perception on the irregular foreign object detection result, performing foreign object approximate distance estimation according to the foreign object state change trend perception result, performing hierarchical warning based on the foreign object approximate distance estimation result, and obtaining a foreign object hierarchical warning result.

2. The method according to claim 1, characterized in that, The method comprises the following steps: performing augmentation and enhancement processing on the video stream image data to obtain enhanced video stream image data, and performing irregular foreign object target feature extraction on the enhanced video stream image data based on an irregular convolution kernel of deformable convolution to obtain an irregular foreign object target feature map; performing feature matching on the irregular foreign object target feature map using a hybrid domain attention mechanism to obtain a feature matching result; performing semantic segmentation-based feature extraction on the feature matching result based on a hollow spatial convolution to obtain dense feature representation, performing feature enhancement on the dense feature representation based on an attention mechanism, and obtaining an irregular foreign object detection result.

3. The method according to claim 2, characterized in that, The method comprises the following steps: performing data augmentation processing on the video stream image data to obtain augmented video stream image data, and performing image aliasing-based data enhancement processing on the augmented video stream image data to obtain enhanced video stream image data; inputting the enhanced video stream image data into a deep convolutional neural network of deformable convolution, and performing preliminary feature extraction using an irregular convolution kernel to obtain an irregular foreign object preliminary feature map; calculating offset amounts in coordinate directions of each sampling point of the irregular foreign object preliminary feature map using a bias layer, and generating a bias matrix according to the offset amount calculation result; updating and optimizing the irregular foreign object preliminary feature map using an interpolation algorithm and a back propagation algorithm based on the sampling point offset amounts in the bias matrix to obtain an irregular foreign object target feature map.

4. The method of claim 2, wherein the method further comprises: The method comprises the following steps: inputting the irregular foreign object target feature map into a channel domain attention mechanism, and performing parallel operations of maximum pooling and average pooling based on the width and height of the irregular foreign object target feature map to obtain two extruded foreign object target feature maps; performing addition operations on the two extruded foreign object target feature maps using a multi-layer perception, and generating normalized channel attention features through a sigmiod activation function; performing multiplication operations on the channel attention features and the irregular foreign object target feature map to obtain an input feature map of a spatial domain attention mechanism; and performing feature matching on the input feature map of the spatial domain attention mechanism using the spatial domain attention mechanism to obtain a feature matching result. performing a serial operation of global max-pooling and average-pooling on the input feature map to obtain two serial operation feature maps, splicing the two serial operation feature maps, and generating spatial dimension attention features by using convolution and sigmoid activation functions; performing channel-by-channel weighting on the spatial dimension attention features to the irregular foreign object target feature map by a multiplication operation to obtain a feature matching result.

5. The method of claim 2, wherein the method further comprises: determining whether the object is a bird or a bird's nest; and providing an alarm if the object is a bird or a bird's nest. The feature matching result is subjected to semantic segmentation-based feature extraction based on the cavity spatial convolution to obtain a dense feature representation, and the dense feature representation is subjected to attention mechanism-based feature enhancement to obtain an irregular foreign object detection result, including: performing three downsampling operations on the feature matching result to obtain a low-dimensional feature map, halving the number of channels of the low-dimensional feature map and dividing it into two parallel branches; each parallel branch uses a depth separable convolution instead of a standard convolution, and uses a cavity convolution with different cavity rates in the convolution layer of each parallel branch to obtain a dense feature representation that fuses multi-scale context information; the dense feature representation is subjected to channel and spatial information expression enhancement by channel attention and spatial attention to obtain an attention branch feature, and a global average pooling branch is introduced to integrate global context information to obtain a global context branch feature; the attention branch feature and the global context branch feature are fused, and the resolution of the feature matching result is matched by an upsampling operation to obtain an irregular foreign object detection result.

6. The method of claim 1, wherein the method further comprises: The irregular foreign object detection result is subjected to foreign object state change trend perception, and foreign object approximate distance estimation is performed according to the foreign object state change trend perception result, and hierarchical warning is performed based on the foreign object approximate distance estimation result to obtain a foreign object hierarchical warning result, including: performing feature fusion processing on the irregular foreign object detection result based on the TransT tracking algorithm to obtain a fusion feature response map, and performing foreign object target tracking and positioning processing on the fusion feature response map to obtain an irregular foreign object motion trajectory; performing foreign object state change trend perception based on time series data on the irregular foreign object motion trajectory by using a change trend predictor to obtain a foreign object state change trend perception result; performing foreign object approximate distance estimation based on semantic segmentation according to the foreign object state change trend perception result to obtain a foreign object approximate distance estimation result; combining the foreign object approximate distance estimation result with a preset warning threshold rule to perform foreign object hierarchical warning to obtain a foreign object hierarchical warning result.

7. The method of claim 6, wherein the method further comprises: determining whether the object is a bird or a bird's nest; and providing an alarm if the object is a bird or a bird's nest. The irregular foreign object detection result is subjected to feature fusion processing based on the TransT tracking algorithm to obtain a fusion feature response map, and foreign object target tracking and positioning processing is performed on the fusion feature response map to obtain an irregular foreign object motion trajectory, including: performing feature extraction on the irregular foreign object detection result by using a TransT tracking algorithm based on a twin network structure to obtain template features and search region features; inputting the template features and the search region features into a feature fusion network, and integrating different position information based on multi-head attention mechanism-based context enhancement and cross feature enhancement to obtain an integration result; encoding the spatial position of the pixels of the integration result by using a sine function to obtain a fusion feature response map; The fusion feature response map is input into the classification branch and the regression branch respectively, and classification output and regression output of the current frame are generated, and the non-regular foreign matter motion trajectory is obtained by screening the anchor frame with the highest foreground classification score by the tracker.

8. The method of claim 6, wherein the method further comprises: The non-regular foreign matter motion trajectory is subjected to foreign matter state change trend perception based on time series data by the change trend predictor, and a foreign matter state change trend perception result is obtained, which includes: When the non-regular foreign matter motion trajectory is in the first frame, the change trend predictor based on Kalman filtering is initialized by the tracker, and an initial state vector is obtained. The state variable is subjected to prior estimation by a state prediction equation, and a prior estimation value is obtained, the prior estimation value of the change trend predictor is corrected based on the tracking frame of the tracker, and the change trend predictor parameters are iteratively updated, and an updated change trend predictor is obtained. The non-regular foreign matter motion trajectory is predicted and updated based on the updated change trend predictor, and a foreign matter state change trend perception result is obtained.

9. The method of claim 6, wherein the method further comprises: determining whether the object is a bird; and determining whether the object is a bird in flight. The foreign matter approximate distance estimation based on semantic segmentation is implemented according to the foreign matter state change trend perception result, and a foreign matter approximate distance estimation result is obtained, which includes: The identified object of semantic segmentation is determined according to the foreign matter state change trend perception result, and pixel-level classification recognition is implemented based on the identified object, and a pixel-level classification recognition result is obtained. The spatial relationship between the two is judged based on the pixel-level classification recognition result, and the spatial relationship judgment result is normalized, and a foreign matter approximate distance estimation result is obtained.

10. The method of claim 6, wherein the method further comprises: determining whether the object is a bird; and if the object is a bird, determining whether the bird is in the transmission path. The hierarchical determination rule of the hierarchical alarm is: When the foreign matter approximate distance estimation result is greater than or equal to 0.3, it is classified as a first-level alarm; When the foreign matter approximate distance estimation result is less than or equal to 0.1, it is classified as a second-level alarm; When the foreign matter approximate distance estimation result is greater than 0.1 and less than 0.3, the alarm level is determined again in combination with the foreign matter type and the weather condition.

Images