Insulator infrared defect diagnosis method based on drone inspection in night environment

Abstract

The present invention discloses an infrared defect diagnosis method for insulators based on drone inspection in a night environment, which belongs to the technical field of electrical equipment safety inspection. A drone is used to perform three-dimensional fixed points on the inspected electrical equipment, and waypoint flight is performed based on the fixed points to obtain images. The image is automatically identified for defects of insulators using an infrared temperature method based on AI technology. The infrared defect diagnosis method for insulators based on drone inspection in a night environment can perform safety inspections of power equipment without installing detection devices in advance by means of drone equipment inspection. At the same time, the power device defect diagnosis technology based on AI technology solves the gap in automatic identification of power insulator defects at night.

Description

Technical Field

[0001] The present invention belongs to the technical field of safety inspection of electrical equipment, and specifically relates to an infrared defect diagnosis method for insulators based on drone inspection in a night environment. Background Art

[0002] With the rapid development of the economy, the whole society has higher requirements for the safety and reliability of the power system. Therefore, the safety and reliability of electrical equipment operation need to be further improved. Infrared thermal imaging temperature measurement technology has the advantages of non-contact, no interference from high-voltage electromagnetic fields, and high efficiency and intuitiveness. It is widely used in regular inspections of power grids to detect abnormal heating defects in a timely manner. At present, faced with a large number of infrared images, inspectors usually perform manual identification and analysis based on work experience. This diagnostic method is not only time-consuming and labor-intensive, but may also result in missed judgments or misjudgments due to over-reliance on manual experience. Therefore, the power device defect diagnosis technology based on AI technology has important engineering practical value.

[0003] In the prior art, most of the methods used to identify defects in power insulators require manual operation, and there is no automated equipment. Even if there is automated equipment, some devices need to be installed in advance. For example:

[0004] 1) Leakage current method: A drainage card and a current sensor are installed on the insulator. It is mainly used to detect the contamination of the insulator. If it is broken, it will not be possible to detect it.

[0005] 2) Ultraviolet imaging method: This needs to be carried out under high humidity conditions, and personnel are required to carry equipment to the top of the tower for identification.

[0006] 3) Infrared temperature measurement method: It also requires personnel to operate it themselves. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides an insulator infrared defect diagnosis method based on drone inspection in a night environment to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: an insulator infrared defect diagnosis method based on drone inspection in a night environment, comprising the following steps:

[0009] S1. Use a drone to perform three-dimensional positioning of the inspected electrical equipment, and perform waypoint flight based on the positioning to obtain images. The drone is equipped with an infrared imager, a high-definition camera, and an IMU navigation device;

[0010] S2. Use the backbone feature extraction network to fragment the acquired image, double the number of channels and half the image dimension. Combine the attention mechanism with the variational inference method to effectively extract and understand multimodal fragmented information and improve the accuracy of automatic image annotation.

[0011] S3, based on the feature fusion module, the features extracted from the backbone features are fused and processed for secondary extraction. The pre-trained ResNet backbone network is used for feature extraction. The multi-scale convolution uses a combination of 1×1 convolution, 3×3 convolution, 5×5 convolution and 7×7 convolution. The instance-level concat operation is performed to merge the feature maps and capture the global background information of the image. The output secondary features are used for classification or regression tasks.

[0012] S4, decoding the secondary features outputted in step S3 to obtain a detection model;

[0013] S5, quantizing and compressing the detection model generated in step S4, wherein the quantization and compression is performed based on the Bayesian method in two ways: uniform quantization or non-uniform quantization;

[0014] S6. Automatically annotate the defect images for training based on semi-automatic annotation technology to obtain a rough label of the defect image, and then fine-tune according to the obtained rough label to form an infrared inspection defect recognition target detection training set, and the training set is used for detection model training;

[0015] S7. Use the integrated edge computing module to accelerate the deep learning of the detection model and form an infrared model for target detection;

[0016] S8, the infrared model of target detection is segmented based on the Mean-Shift algorithm (mean shift algorithm) to obtain a picture without background. The Mean Shift algorithm process is: calculate the mean shift vector m of each sample h (x), for each sample point m h (x) is translated, that is: x i =x i +m h (x i ), repeat the above steps until the sample points converge, that is: m h (x) = 0, samples that converge to the same point are considered to be members of the same cluster;

[0017] S9. Adding an AI online recognition module to the infrared detection method to cut the image and obtain the defect image;

[0018] S10, performing filtering and cutting processing on the obtained defect image to segment out abnormal points with defect problems in the device;

[0019] S11. Save the abnormal point image and generate a corresponding defect report.

[0020] To further optimize the technical solution, in S2, in order to avoid feature loss of the target image after multiple convolutions, the last three layers of features of the backbone feature extraction network are extracted for subsequent processing.

[0021] To further optimize the technical solution, in S3, the feature fusion module performs secondary extraction processing on the features based on a multi-scale neural network structure, and the features of different dimensions in step S2 are fused and extracted through the feature fusion module.

[0022] To further optimize the technical solution, in S5, quantization compression is used to reduce model parameters and improve detection speed, and a model distillation method is used to improve the accuracy of the model.

[0023] To further optimize the technical solution, in S6, defect images for training are extracted based on the infrared power device database, and the extracted images are subjected to data augmentation processing.

[0024] To further optimize the technical solution, in S7, the target detection infrared model needs to be preprocessed to eliminate noise contained in the infrared image.

[0025] To further optimize the technical solution, in S9, the infrared detection method includes a surface temperature determination method, a horizontal comparison determination method, a vertical comparison determination method, a thermal image feature determination method and a relative temperature difference determination method.

[0026] To further optimize the technical solution, in S9, the AI ​​online recognition module realizes the display type and position of the target in the real-time picture based on a multi-scale pyramid neural network.

[0027] Compared with the prior art, the present invention provides an insulator infrared defect diagnosis method based on drone inspection in a night environment, which has the following beneficial effects:

[0028] This infrared defect diagnosis method for insulators based on drone inspections in nighttime environments can conduct safety inspections of power equipment through drone equipment inspections without the need to install detection devices in advance. At the same time, the power device defect diagnosis technology based on AI technology solves the gap in automated identification of power insulator defects at night. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a schematic diagram of the connection structure of an AI online recognition module in an insulator infrared defect diagnosis method based on drone inspection in a night environment proposed by the present invention;

[0030] Figure 2This is a diagram showing the effect of filtering and cutting processing in an insulator infrared defect diagnosis method based on drone inspection in a night environment proposed by the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Embodiment 1:

[0033] An insulator infrared defect diagnosis method based on drone inspection in a night environment includes the following steps:

[0034] S1. Use a drone to perform three-dimensional positioning of the inspected electrical equipment, and perform waypoint flight based on the positioning to obtain images. The drone is equipped with an infrared imager, a high-definition camera, and an IMU navigation device;

[0035] S2. Use the backbone feature extraction network to fragment the acquired image, double the number of channels and half the image dimension. Combine the attention mechanism with the variational inference method to effectively extract and understand multimodal fragmented information and improve the accuracy of automatic image annotation.

[0036] S3, based on the feature fusion module, the features extracted from the backbone features are fused and processed for secondary extraction. The pre-trained ResNet backbone network is used for feature extraction. The multi-scale convolution uses a combination of 1×1 convolution, 3×3 convolution, 5×5 convolution and 7×7 convolution. The instance-level concat operation is performed to merge the feature maps and capture the global background information of the image. The output secondary features are used for classification or regression tasks.

[0037] S4, decoding the secondary features outputted in step S3 to obtain a detection model;

[0038] S5, performing quantization compression processing on the detection model generated in step S4, wherein the quantization compression is performed based on the Bayesian method using two methods of uniform quantization or non-uniform quantization;

[0039] S6. Automatically annotate the defect images for training based on semi-automatic annotation technology to obtain a rough label of the defect image, and then fine-tune according to the obtained rough label to form an infrared inspection defect recognition target detection training set, and the training set is used for detection model training;

[0040] S7. Use the integrated edge computing module to accelerate the deep learning of the detection model and form an infrared model for target detection;

[0041] S8, the infrared model of target detection is segmented based on the Mean-Shift algorithm to obtain a background-free image. The Mean Shift algorithm process is as follows: calculate the mean shift vector m of each sample h (x), for each sample point m h (x) is translated, that is: x i =x i +m h (x i ), repeat the above steps until the sample points converge, that is: m h (x) = 0, samples that converge to the same point are considered to be members of the same cluster;

[0042] S9. Adding an AI online recognition module to the infrared detection method to cut the image and obtain the defect image;

[0043] S10, performing filtering and cutting processing on the obtained defect image to segment out abnormal points with defect problems in the device;

[0044] S11. Save the abnormal point image and generate a corresponding defect report.

[0045] Specifically, in S2, in order to avoid feature loss of the target image after multiple convolutions, the last three layers of features of the backbone feature extraction network are extracted for subsequent processing.

[0046] Specifically, in S3, the feature fusion module performs secondary extraction processing on the features based on a multi-scale neural network structure, and the features of different dimensions in step S2 are fused and extracted through the feature fusion module.

[0047] Specifically, in S5, quantization compression is used to reduce model parameters and improve detection speed, and a model distillation method is used to improve the accuracy of the model.

[0048] Specifically, in S6, defective images for training are extracted based on the infrared power device database, and the extracted images are subjected to data augmentation processing.

[0049] Specifically, in S7, the target detection infrared model needs to perform preprocessing for eliminating noise contained in the infrared image.

[0050] Specifically, in S9, the AI ​​online recognition module realizes the display type and position of the target in the real-time picture based on a multi-scale pyramid neural network.

[0051] Embodiment 2:

[0052] An insulator infrared defect diagnosis method based on drone inspection in a night environment includes the following steps:

[0053] S1. Use a drone to perform three-dimensional positioning of the inspected electrical equipment, and perform waypoint flight based on the positioning to acquire images. The drone is equipped with an infrared imager, a high-definition camera, and an IMU navigation device.

[0054] S2. Use the backbone feature extraction network to fragment the acquired image, reduce the number of channels to twice the original number, and reduce the image dimension to 1 / 2 of the original number. Combine the attention mechanism with the variational inference method to effectively extract and understand multimodal fragmented information and improve the accuracy of automatic image annotation. In order to avoid the loss of features of the target image after multiple convolutions, the last three layers of features of the backbone feature extraction network are extracted for subsequent processing.

[0055] S3. Based on the feature fusion module, the features extracted from the backbone features are fused and secondary extracted. The pre-trained ResNet backbone network is used for feature extraction. The multi-scale convolution uses a combination of 1×1 convolution, 3×3 convolution, 5×5 convolution and 7×7 convolution. The instance-level concat operation is performed to merge the feature maps and capture the global background information of the image. The output secondary features are used for classification or regression tasks. The feature fusion module performs secondary extraction on the features based on the multi-scale neural network structure, and the features of different dimensions in step S2 are fused and extracted through the feature fusion module.

[0056] S4. Decode the secondary features output from step S3 to obtain a detection model.

[0057] S5. The detection model generated in step S4 is quantized and compressed. The quantization compression is performed based on the Bayesian method using uniform quantization or non-uniform quantization to reduce model parameters and increase detection speed. The model distillation method is used to improve the accuracy of the model.

[0058] S6. Automatically annotate the defect images used for training based on semi-automatic annotation technology to obtain a rough label of the defect image, and then fine-tune it according to the obtained rough label to form an infrared inspection defect recognition target detection training set, and perform offline model training on the training set.

[0059] Among them, Figure 1 As shown, defective images for training are extracted based on the infrared power device database, and the extracted images are subjected to data augmentation processing. The above process belongs to the preparation process of the data set.

[0060] S7. Use the integrated edge computing module to accelerate the deep learning of the detection model to form a target detection infrared model. The target detection infrared model requires preprocessing to eliminate the noise contained in the infrared image.

[0061] S8, the infrared model of target detection is segmented based on the Mean-Shift algorithm (mean shift algorithm) to obtain a picture without background. The Mean Shift algorithm process is: calculate the mean shift vector m of each sample h (x), for each sample point m h (x) is translated, that is: x i =x i +m h (x i ), repeat the above steps until the sample points converge, that is: m h (x) = 0, the samples that converge to the same point are considered to be members of the same cluster.

[0062] S9. Add an AI online recognition module to the infrared detection method to cut the image to obtain the defect image. The infrared detection method includes surface temperature judgment method, horizontal comparison judgment method, vertical comparison judgment method, thermal image feature judgment method and relative temperature difference judgment method.

[0063] Among them, Figure 1 As shown, the video stream is input into the AI ​​online recognition module. The AI ​​online recognition module realizes the display of the type and position of the target in the real-time picture based on the multi-scale pyramid neural network, forms a result schematic diagram based on AI technology, and extracts the picture of the power equipment from the result schematic diagram.

[0064] S10, performing filtering and cutting processing on the obtained defective image to segment out abnormal points with defective problems in the device.

[0065] Among them, Figure 2 As shown, after the filtered infrared image is cut, a defect image without background and only highlighting the defects is generated.

[0066] S11. Save the abnormal point image and generate a corresponding defect report.

[0067] The beneficial effects of the present invention are: the infrared defect diagnosis method for insulators based on drone inspections in nighttime environments can perform safety inspections of power equipment through drone equipment inspections without the need to install detection devices in advance. At the same time, the power device defect diagnosis technology based on AI technology solves the gap in automated identification of power insulator defects at night.

[0068] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An infrared defect diagnosis method for insulators based on drone inspection in a night environment, characterized in that: The following steps are involved: S1. Use a drone to perform three-dimensional positioning of the inspected electrical equipment, and perform waypoint flight based on the positioning to obtain images. The drone is equipped with an infrared imager, a high-definition camera, and an IMU navigation device; S2. Use the backbone feature extraction network to fragment the acquired image, double the number of channels and half the image dimension. Combine the attention mechanism with the variational inference method to effectively extract and understand multimodal fragmented information and improve the accuracy of automatic image annotation. S3, based on the feature fusion module, the features extracted from the backbone features are fused and processed for secondary extraction. The pre-trained ResNet backbone network is used for feature extraction. The multi-scale convolution uses a combination of 1×1 convolution, 3×3 convolution, 5×5 convolution and 7×7 convolution. The instance-level concat operation is performed to merge the feature maps and capture the global background information of the image. The output secondary features are used for classification or regression tasks. S4, decoding the secondary features outputted in step S3 to obtain a detection model; S5, performing quantization compression processing on the detection model generated in step S4, wherein the quantization compression is performed based on the Bayesian method using two methods of uniform quantization or non-uniform quantization; S6. Automatically annotate the defect images for training based on semi-automatic annotation technology to obtain a rough label of the defect image, and then fine-tune according to the obtained rough label to form an infrared inspection defect recognition target detection training set, and the training set is used for detection model training; S7. Use the integrated edge computing module to accelerate the deep learning of the detection model and form an infrared model for target detection; S8, the infrared model of target detection is segmented based on the Mean Shift algorithm to obtain a background-free image. The MeanShift algorithm process is as follows: Calculate the mean shift vector m of each sample point h (x), for each sample point m h (x) is translated, that is: x i =x i +m h (x i ), repeat the above steps until the sample points converge, that is: m h (x) = 0, the sample points that converge to the same point are considered to be members of the same cluster; S9. Adding an AI online recognition module to the infrared detection method to cut the image and obtain the defect image; S10, performing filtering and cutting processing on the obtained defect image to segment out abnormal points with defect problems in the device; S11. Save the abnormal point image and generate a corresponding defect report.

2. The infrared defect diagnosis method for insulators based on drone inspection in a night environment according to claim 1 is characterized in that: In S2, in order to avoid feature loss of the target image after multiple convolutions, the last three layers of features of the backbone feature extraction network are extracted for subsequent processing.

3. The insulator infrared defect diagnosis method based on drone inspection in a night environment according to claim 1 is characterized in that: In S3, the feature fusion module performs secondary extraction processing on the features based on a multi-scale neural network structure, and the features of different dimensions in step S2 are fused and extracted through the feature fusion module.

4. The insulator infrared defect diagnosis method based on drone inspection in a night environment according to claim 1 is characterized in that: In S5, quantization compression is used to reduce model parameters and improve detection speed, and a model distillation method is used to improve the accuracy of the model.

5. The insulator infrared defect diagnosis method based on drone inspection in nighttime environment according to claim 1 is characterized in that: In S6, defective images for training are extracted based on the infrared power device database, and the extracted images are subjected to data augmentation processing.

6. The method for diagnosing infrared defects of insulators based on drone inspection in a nighttime environment according to claim 1 is characterized in that: In S7, the target detection infrared model needs to perform preprocessing for eliminating noise contained in the infrared image.

7. The method for diagnosing infrared defects of insulators based on drone inspection in a nighttime environment according to claim 1 is characterized in that: In the above S9, the infrared detection method includes a surface temperature determination method, a horizontal comparison determination method, a vertical comparison determination method, a thermal image feature determination method and a relative temperature difference determination method.

8. The insulator infrared defect diagnosis method based on drone inspection in nighttime environment according to claim 1 is characterized in that: In S9, the AI ​​online recognition module realizes displaying the type and position of the target in the real-time picture based on a multi-scale pyramid neural network.

Images