A method and device for detecting transmission line insulator faults based on artificial intelligence

Abstract

The present invention provides a method for detecting insulator faults of transmission lines based on artificial intelligence, comprising the following steps: obtaining historical insulator images of several transmission lines through a drone, and marking the status of the insulators in each of the images; preprocessing the images and using them as training data sets; the preprocessing includes Gaussian blurring, Gaussian noise addition, and image sharpening; constructing a fault detection model based on YOLO, inputting the training data set into the model for training, and updating the hyperparameters of the model; inputting the transmission line insulator images to be detected into the trained fault detection model, and obtaining the corresponding insulator status detection results. The present invention detects insulator faults based on artificial intelligence technology, effectively solves the technical problem of many types of insulator defects in the prior art, realizes the automatic detection of insulator faults, and significantly improves the efficiency and accuracy of insulator fault detection.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based transmission line insulator fault detection method, device and computer-readable storage medium. Background Art

[0002] In transmission lines, insulators are necessary equipment with double insulation because they mechanically fix the wires and enable the transmission lines to operate reliably and uninterruptedly. However, due to external environmental factors, insulators often suffer local damage, and any faults or defects on the insulators will lead to unstable power supply and endanger the safety of the transmission line. Therefore, the status monitoring of insulators is an important part of the inspection of transmission lines. Timely monitoring and intelligent detection of defects on transmission line insulators will reduce the occurrence of power outages and other faults.

[0003] Early inspection methods usually adopt patrol inspection, such as manual patrol inspection, which is costly and inefficient, or mechanical equipment patrol inspection, such as helicopter inspection. However, this method is inefficient and requires specialized skills to operate the corresponding mechanical equipment, and has poor universality in practice. With the rapid development of aviation technology, drones have attracted widespread attention due to their fast speed, work efficiency and reliability in inspecting power equipment. Drones with airborne cameras can capture aerial images of insulators and power lines, which can be used as photographic records for the operation and maintenance of transmission line insulators. Fault detection of insulators by analyzing insulator images has become the mainstream technology at present. It is mainly realized by digital image processing technology, such as image segmentation, filter and wavelet transform. However, these traditional methods can only detect fixed textures, and their generalization ability is weak, and the detection speed and detection accuracy are not high. At the same time, due to the large number of insulator faults, different image analysis algorithms need to be developed for each fault, which is relatively inefficient. Therefore, the technical problem of the prior art is that the detection accuracy and detection efficiency are not high when performing insulator fault detection. Summary of the invention

[0004] The purpose of the present invention is to provide a transmission line insulator fault detection method based on artificial intelligence, which can solve the technical problems existing in the prior art that the insulator fault detection accuracy and detection efficiency are not high.

[0005] A first aspect of the present invention provides a method for detecting transmission line insulator faults based on artificial intelligence, comprising the following steps:

[0006] S1: Obtain historical insulator images of several transmission lines through a drone, and annotate the status of the insulators in each of the images;

[0007] S2: preprocessing the image and using it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise and image sharpening;

[0008] S3: construct a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model;

[0009] S4: Input the transmission line insulator image to be detected into the trained fault detection model, and obtain the corresponding insulator state detection result.

[0010] As a further improvement of the present invention, the YOLO-based fault detection model includes: a backbone network layer (Backbone layer), a neck layer (Neck layer), and a head layer (Head layer); the model inputs an image into the Backbone layer; the Backbone layer includes a Focus layer, a CFF module, a first CSP, a first CBL, a second CSP, a second CBL, an SPP, and a third CSP connected in sequence; the Neck layer includes a first concat to a fourth concat, a first upsampling, a second upsampling, a third CBL to a tenth CBL, and a fourth CSP to an eighth CSP; the Head layer includes a first feature map P1 to a fourth feature map P4; the first C The output of CSP is input to the first concat, and the output of the second CSP is input to the second concat; the fourth CBL, the second upsampling, the second concat, the fourth CSP, the third CBL, the first upsampling, the first concat, the fifth CSP, the seventh CBL, and P1 are connected in sequence; the third CBL, the third concat, the seventh CSP, the eighth CBL, and P2 are connected in sequence; the fifth CSP, the fifth CBL, the third concat, the sixth CSP, the sixth CBL, and the fourth concat are connected in sequence; the sixth CSP, the ninth CBL, and P3 are connected in sequence; the third CSP, the fourth CBL, the fourth concat, the eighth CSP, the tenth CBL, and P4 are connected in sequence.

[0011] As a further improvement of the present invention, the CFF module includes CBL1~CBL4, a maximum pooling layer and concat; the input feature map is sent to CBL1, and its output is respectively input to CBL2 and the maximum pooling layer, the output of CBL2 is input to CBL3, the output of CBL3 and the maximum pooling layer is input to concat, and the output of concat is input to CBL4.

[0012] As a further improvement of the present invention, the first to third CSPs are CSP1_X, and the fourth to eighth CSPs adopt improved CSPs; the improved CSPs include: CBL5 to CBL10, Ghost bottleneck layer, depth convolution, point convolution, average pooling, Sigmoid function, and concat; the input feature maps are respectively input into CBL5 and CBL6; CBL5 is connected to the Ghost bottleneck layer, and the Ghost bottleneck layer outputs two branches, the first branch includes CBL7, depth convolution, and point convolution connected in sequence, and the second branch includes average pooling, CBL8, and Sigmoid function connected in sequence, the output of the Sigmoid function is cross-multiplied with the output of the Ghost bottleneck layer, and the result is input into CBL9, the outputs of CBL6, point convolution, and CBL9 are input into concat, and the output result of concat is input into CBL10.

[0013] As a further improvement of the present invention, the error function used in the training process is: Where IoU is the intersection ratio of the real frame and the predicted frame, d1 is the distance between the center points of the predicted frame and the real frame, c is the minimum diagonal distance between the real frame and the predicted frame, d2 is the distance between the upper left corner coordinates of the predicted frame and the real frame, and r is the ratio of the width of the predicted frame to the real frame.

[0014] As a further improvement of the present invention: the model uses an improved firefly algorithm to update the hyperparameters of the model, and maps the hyperparameters to the positions of fireflies. The algorithm includes:

[0015] S61: Initialize algorithm parameters, including population size, current iteration number G, maximum iteration number G max , objective function f = ADIOU;

[0016] S62: Randomly initialize the positions of all individuals in the population;

[0017] S63: Start iteration: Calculate the Euclidean distance r between individual P(i,G) and individual P(j,G) i,j , P(i,G) is the position parameter of individual i in the Gth iteration; three individuals d, e, and f are randomly selected from the population, denoted as P(d,G), P(e,G), and P(f,G), and they do not overlap with individual i;

[0018] S64: Calculate the objective function values ​​of all individuals according to the objective function and sort them, and record the individual b with the smallest current objective function value as P(b,G);

[0019] S65: Calculate the attraction between individuals i and j: where f mean、f best are the mean and minimum values ​​of the objective function values ​​of all individuals in the current iteration process, β0 is the maximum attraction between individuals, γ is the absorption coefficient, and the scale factor It follows a normal distribution with a mean of 0.5 and a variance of 0.3;

[0020] S66: Calculate the adaptive threshold Δ=1-|f mean -f best | / f worst -f best |, where f worst It is the maximum value of the objective function of all individuals in the current iteration process;

[0021] S67: If Δ≥0.5, update the position of the firefly using the following formula: P(i,G+1) is the position of individual i after the G+1th iteration, β(i,d) is the attraction between individuals i and d, is the scale factor, which follows a normal distribution with a mean of 0.5 and a variance of 0.3; otherwise, the position is updated using the following formula:

[0022] S68: Determine whether the maximum number of iterations has been reached. If so, output the individual position with the smallest objective function value as the result; otherwise, proceed to S63.

[0023] The second aspect of the present invention provides a transmission line fault detection device based on artificial intelligence, characterized in that it includes a data acquisition module, a data preprocessing module, a model training module, and a fault detection module, wherein:

[0024] A data acquisition module, used to acquire historical insulator images of a number of transmission lines through a drone, and to mark the status of the insulators in each of the images;

[0025] A data preprocessing module, used to preprocess the image and use it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise reduction and image sharpening;

[0026] A model training module, used to build a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model;

[0027] The fault detection module is used to input the transmission line insulator image to be detected into the trained fault detection model and obtain the corresponding insulator state detection result.

[0028] A third aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the artificial intelligence-based transmission line fault detection method as described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The present invention realizes insulator fault detection based on artificial intelligence technology, effectively solves the technical problem of many types of insulator defects in the prior art, realizes automatic detection of insulator faults, and significantly improves the efficiency and accuracy of insulator fault detection;

[0031] (2) The present invention improves the YOLO model, adds a CFF module in the Backbone layer, and improves the CSP module. The above method can extract more effective features, increase the complexity of the extracted features, ensure the effectiveness of the extracted features, and only slightly increase the computational complexity while improving the detection accuracy;

[0032] (3) The present invention uses an improved firefly algorithm to optimize the hyperparameters of the model, thereby improving the convergence speed of the algorithm and avoiding falling into a local optimal value, thereby obtaining the optimal hyperparameters of the model, which can further improve the accuracy of insulator detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of a transmission line insulator fault detection method based on artificial intelligence provided by an embodiment of the present invention.

[0034] Figure 2 An architecture of a YOLO fault detection model provided by an embodiment of the present invention.

[0035] Figure 3 An architecture of a Focus layer in a YOLO fault detection model provided by an embodiment of the present invention.

[0036] Figure 4 The architecture of CSP1_X provided in the embodiment of the present invention.

[0037] Figure 5 An improved CSP architecture is provided in an embodiment of the present invention.

[0038] Figure 6 The architecture of the Ghost bottleneck layer provided by the embodiment of the present invention.

[0039] Figure 7 A CFF module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] Figure 1 A flow chart of a transmission line fault detection method based on artificial intelligence provided by an embodiment of the present invention includes the following steps:

[0041] S1: Obtain historical insulator images of several transmission lines through a drone, and annotate the status of the insulators in each of the images;

[0042] S2: preprocessing the image and using it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise and image sharpening;

[0043] S3: construct a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model;

[0044] S4: Input the transmission line insulator image to be detected into the trained fault detection model, and obtain the corresponding insulator state detection result.

[0045] The present invention inputs the insulator image into the YOLO fault detection model for learning, updates the hyperparameters of the fault detection model, and uses the trained model for fault detection. The method effectively solves the technical problem of many types of insulator defects in the prior art based on artificial intelligence technology, realizes the automatic detection of insulator faults, and significantly improves the efficiency and accuracy of insulator fault detection.

[0046] As a further improvement of the present invention, the YOLO-based fault detection model includes: a backbone network layer (Backbone layer), a neck layer (Neck layer), and a head layer (Head layer); the model inputs an image into the Backbone layer; the Backbone layer includes a Focus layer, a CFF module, a first CSP, a first CBL, a second CSP, a second CBL, an SPP, and a third CSP connected in sequence; the Neck layer includes a first concat to a fourth concat, a first upsampling, a second upsampling, a third CBL to a tenth CBL, and a fourth CSP to an eighth CSP; the Head layer includes a first feature map P1 to a fourth feature map P4; the first C The output of CSP is input to the first concat, and the output of the second CSP is input to the second concat; the fourth CBL, the second upsampling, the second concat, the fourth CSP, the third CBL, the first upsampling, the first concat, the fifth CSP, the seventh CBL, and P1 are connected in sequence; the third CBL, the third concat, the seventh CSP, the eighth CBL, and P2 are connected in sequence; the fifth CSP, the fifth CBL, the third concat, the sixth CSP, the sixth CBL, and the fourth concat are connected in sequence; the sixth CSP, the ninth CBL, and P3 are connected in sequence; the third CSP, the fourth CBL, the fourth concat, the eighth CSP, the tenth CBL, and P4 are connected in sequence.

[0047] The YOLO model is a regression method based on deep learning and is widely used in image classification. Figure 2 A YOLO model is provided in an embodiment of the present invention. The Focus layer is used to reduce the computational cost and improve the speed of the model. The Focus layer slices the feature map and integrates the width and height information into the channel. Figure 3 This is a schematic diagram of the Focus module architecture provided by an embodiment of the present invention, including 4 slice modules, a concat layer and a CBL. The output feature maps of the traditional YOLO model only include three types: P1 (80*80*16), P2 (40*40*16) and P3 (20*20*16). The present invention adds an additional P4 output block, whose feature map size is 10×10×16, which can enhance the model's ability to detect small-sized targets while ensuring detection efficiency.

[0048] CBL consists of a convolutional layer (Conv), a batch normalization (BN layer) and a Leaky Relu activation function layer. CBL is the abbreviation of the above three components.

[0049] CSP (Cross Stage Partial), also known as cross-stage network, includes two structures, CSP1_X and CSP2_X, which are used for Backbone layer and Neck layer respectively. The first to third CSPs in the present invention belong to CSP1_X. See Figure 4 As shown, X is the number of CBLs in the residual structure, and the values ​​of X for the first to third CSPs are 1, 3, and 3 respectively. The difference between CSP2_X and CSP1_X is that the former uses 2X CBLs to replace the residual structure.

[0050] In order to further improve the effectiveness of the extracted features, the present invention improves the CSP2_X structure. The third CSP to the eighth CSP in the embodiment of the present invention all adopt the improved CSP structure. Figure 5 The improved CSP structure provided by the embodiment of the present invention is Figure 6 The Ghost bottleneck layer used in the improved CSP structure. The improved CSP includes: CBL5~CBL10, Ghost bottleneck layer, depth convolution, point convolution, average pooling, Sigmoid function, concat; the input feature map is input to CBL5 and CBL6 respectively; CBL5 is connected to the Ghost bottleneck layer, and the Ghost bottleneck layer outputs two branches. The first branch includes CBL7, depth convolution, and point convolution connected in sequence, and the second branch includes average pooling, CBL8, and Sigmoid function connected in sequence. The output of the Sigmoid function is cross-multiplied with the output of the Ghost bottleneck layer, and the result is input to CBL9. The outputs of CBL6, point convolution, and CBL9 are input to concat, and the output result of concat is input to CBL10. The improved CSP structure replaces the two ordinary convolution layers in the original CSP structure with the Ghost bottleneck layer, and then the output of this part is divided into two parts, one part extracts more information through the inverted residual structure, and the other part extracts global information through the global average pooling layer to enhance the ability to extract global information. These two parts can avoid data redundancy caused by repeated extraction of the same features. Finally, the outputs of these three parts are concatenated in the channel dimension, and then the number of final output channels is adjusted through a convolutional layer. The improved CSP structure is more complex than the original CSP structure, but due to the use of certain lightweight structures, the amount of calculation has not increased significantly, but the improved CSP structure can extract more feature information and enhance the detection effect of the model.

[0051] The Ghost bottleneck layer is a bottleneck structure composed of Ghost modules. In essence, Ghost Module is used to replace the ordinary convolution in the bottleneck structure. Ghost convolution is to use 1x1 convolution to obtain the necessary feature concentration of input features. 1x1 convolution is used to extract cross-channel features of the feature map we input, compress the channels, and obtain a feature concentration.

[0052] As a further improvement of the present invention, the CFF module includes CBL1, whose outputs are respectively input into CBL2 and the maximum pooling layer, the output of CBL2 is input into CBL3, the output of CBL3 and the maximum pooling layer are input into concat, and the concat output is input into CBL4.

[0053] In order to further improve the accuracy of the model, the present invention designs a CFF module (Complicated Feature Fusion). Figure 6 It further downsamples the input image by using a convolution with a stride of 2 and increases the number of channels, which increases the complexity of the extracted features and ensures the effectiveness of the extracted features. This module only slightly increases the computational complexity while ensuring improved detection accuracy.

[0054] As a further improvement of the present invention, the error function used in the training process is: Where IoU is the intersection ratio of the real frame and the predicted frame, d1 is the distance between the center points of the predicted frame and the real frame, c is the minimum diagonal distance between the real frame and the predicted frame, d2 is the distance between the upper left corner coordinates of the predicted frame and the real frame, and r is the ratio of the width of the predicted frame to the real frame.

[0055] When training the YOLO model, the commonly used error functions are GIoU and DIoU. However, for DIoU, when the distance between the two target boxes remains unchanged, the longer the distance of the center of mass c, the smaller the value of the DIoU loss function, which indicates that the DIoU loss function may achieve the purpose of reducing the loss function by expanding the prediction box. Therefore, the present invention corrects DIoU and improves the DIoU loss function to reduce the magnitude of the change in the DIoU value when the length of the prediction box changes, thereby accelerating the convergence of the model. At the same time, the formula of the present invention introduces the ratio of the width of the prediction box to the true box. When the length of the sides of the prediction box and the true box changes, the problem of the loss value decreasing as the length of the prediction box is expanded or reduced can be solved.

[0056] As a further improvement of the present invention, the model uses an improved firefly algorithm to update the hyperparameters of the model, and maps the hyperparameters to the positions of fireflies. The algorithm includes:

[0057] S61: Initialize algorithm parameters, including population size, current iteration number G, maximum iteration number G max , objective function f = ADIOU;

[0058] S62: Randomly initialize the positions of all individuals in the population;

[0059] S63: Start iteration: Calculate the Euclidean distance r between individual P(i,G) and individual P(j,G) i,j , P(i,G) is the position parameter of individual i in the Gth iteration; three individuals d, e, and f are randomly selected from the population, denoted as P(d,G), P(e,G), and P(f,G), and they do not overlap with individual i;

[0060] S64: Calculate the objective function values ​​of all individuals according to the objective function and sort them, and record the individual b with the smallest current objective function value as P(b,G);

[0061] S65: Calculate the attraction between individuals i and j: where f mean 、f best are the mean and minimum values ​​of the objective function values ​​of all individuals in the current iteration process, β0 is the maximum attraction between individuals, γ is the absorption coefficient, and the scale factor It follows a normal distribution with a mean of 0.5 and a variance of 0.3;

[0062] S66: Calculate the adaptive threshold Δ=1-|f mean -f best | / |f worst -f best |, where f worst It is the maximum value of the objective function of all individuals in the current iteration process;

[0063] S67: If Δ≥0.5, update the position of the firefly using the following formula: P(i,G+1) is the position of individual i after the G+1th iteration, β(i,d) is the attraction between individuals i and d, is the scale factor, which follows a normal distribution with a mean of 0.5 and a variance of 0.3; otherwise, the position is updated using the following formula:

[0064] S68: Determine whether the maximum number of iterations has been reached. If so, output the individual position with the smallest objective function value as the result; otherwise, proceed to S63.

[0065] The firefly algorithm is a widely used optimization algorithm and a heuristic optimization algorithm that simulates the behavior of fireflies in nature. It is similar to genetic algorithms, particle swarm optimization algorithms, etc., and is often used to solve global optimization problems. In the early stages of the evolution process, the distance between the firefly and the global optimal solution is extremely large; therefore, the attraction tends to zero and is almost non-existent. Then, the movement of the firefly is based on a random walk in the search space, so the convergence speed of the algorithm cannot be guaranteed. On the other hand, as the number of generations increases, the distance between the firefly and the global optimal solution becomes extremely small; therefore, the attraction remains unchanged. Therefore, it is necessary to modify the search process of the firefly algorithm to enhance the optimization performance in terms of the balance between global exploration and local development capabilities. The present invention modifies the firefly attraction formula and sets an adaptive threshold to update the position, so that the attraction gradually increases with the growth of the number of generations to enhance the ability to find the global optimal solution, help maintain the diversity within the group, and improve the convergence speed of the algorithm to create a balance between global exploration and local search capabilities.

[0066] The scale factor in the above algorithm uses a normally distributed variable. In order to further improve the algorithm effect, the following method can also be used to generate the scale factor:

[0067] Assume that the mutant individual is V(i,G), and the mutant individual is generated by the following formula:

[0068]

[0069] Defining the crossover rate The test individuals are generated based on the crossover rate as follows:

[0070]

[0071] Where: represents the test individual generated after the crossover operation, rand is a random number between 0 and 1; the crossover rate and the scale factor The update formula is as follows:

[0072]

[0073] In the formula, randn and randc represent normal distribution and Cauchy distribution respectively. and represents the peak location parameter and mean of the Cauchy distribution and the normal distribution, and are stored in S respectively. F,k and S CR,k In , they are updated by the following formula:

[0074]

[0075]

[0076]

[0077] Where: n is S F,k and S CR,k The number of elements in .

[0078] The second aspect of the present invention provides a transmission line fault detection device based on artificial intelligence, characterized in that it includes a data acquisition module, a data preprocessing module, a model training module, and a fault detection module, wherein:

[0079] A data acquisition module, used to acquire historical insulator images of a number of transmission lines through a drone, and to mark the status of the insulators in each of the images;

[0080] A data preprocessing module, used to preprocess the image and use it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise reduction and image sharpening;

[0081] A model training module, used to build a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model;

[0082] The fault detection module is used to input the transmission line insulator image to be detected into the trained fault detection model and obtain the corresponding insulator state detection result.

[0083] A third aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the artificial intelligence-based transmission line fault detection method as described above.

[0084] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0085] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for detecting transmission line insulator faults based on artificial intelligence, comprising the following steps: S1: Obtain historical insulator images of several transmission lines through a drone, and annotate the status of the insulators in each of the images; S2: preprocessing the image and using it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise and image sharpening; S3: construct a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model; S4: inputting the transmission line insulator image to be detected into the trained fault detection model, and obtaining the corresponding insulator state detection result; The YOLO-based fault detection model includes: a backbone network layer (Backbone layer), a neck layer (Neck layer), and a head layer (Head layer); the model inputs an image into the Backbone layer; the Backbone layer includes a Focus layer, a CFF module, a first CSP, a first CBL, a second CSP, a second CBL, an SPP, and a third CSP connected in sequence; the Neck layer includes a first concat to a fourth concat, a first upsampling, a second upsampling, a third CBL to a tenth CBL, and a fourth CSP to an eighth CSP; the Head layer includes a first feature map P1 to a fourth feature map P4; the output of the first CSP is input to the first concat, and the output of the second CSP is input to the second concat; the fourth CBL, the second upsampling, the second concat, the fourth CSP, the first concat, the second CSP, the third CBL, the fourth CSP ... The third CBL, the first upsampling, the first concat, the fifth CSP, the seventh CBL, and P1 are connected in sequence; the third CBL, the third concat, the seventh CSP, the eighth CBL, and P2 are connected in sequence; the fifth CSP, the fifth CBL, the third concat, the sixth CSP, the sixth CBL, and the fourth concat are connected in sequence; the sixth CSP, the ninth CBL, and P3 are connected in sequence; the third CSP, the fourth CBL, the fourth concat, the eighth CSP, the tenth CBL, and P4 are connected in sequence; the CFF module includes CBL1 to CBL4, a maximum pooling layer and concat; the input feature map is sent to CBL1, and its output is respectively input to CBL2 and the maximum pooling layer, the output of CBL2 is input to CBL3, the output of CBL3 and the maximum pooling layer is input to concat, and the output of concat is input to CBL4; The error function used in the training process is: Where IoU is the cross ratio of the real frame and the predicted frame, d1 is the distance between the center points of the predicted frame and the real frame, c is the minimum diagonal distance between the real frame and the predicted frame, d2 is the distance between the upper left corner coordinates of the predicted frame and the real frame, and r is the ratio of the width of the predicted frame to the real frame; The model uses an improved firefly algorithm to update the hyperparameters of the model and maps the hyperparameters to the positions of fireflies. The algorithm includes: S61: Initialize algorithm parameters, including population size, current iteration number G, maximum iteration number G max , objective function f = ADIOU; S62: Randomly initialize the positions of all individuals in the population; S63: Start iteration: Calculate the Euclidean distance r between individual P(i, G) and individual P(j, G) i，j , P(i, G) is the position parameter of individual i in the Gth iteration; three individuals d, e, and f are randomly selected from the population, denoted as P(d, G), P(e, G), and P(f, G), and they do not overlap with individual i; S64: Calculate the objective function values ​​of all individuals according to the objective function and sort them, and record the individual b with the smallest current objective function value as P(b, G); S65: Calculate the attraction between individuals i and j: where f mean 、f best are the mean and minimum values ​​of the objective function values ​​of all individuals in the current iteration process, β0 is the maximum attraction between individuals, γ is the absorption coefficient, and the scale factor F i G It follows a normal distribution with a mean of 0.5 and a variance of 0.3; S66: Calculate adaptive threshold Δ=1-|f mean -f best | / |f worst -f best |, where f worst It is the maximum value of the objective function of all individuals in the current iteration process; S67: If Δ≥0.5, update the position of the firefly using the following formula: P(i,G+1)=(1-β(i,d))P(i,G)+β(i,d)P(i,G)+F i G (P(e, G)-P(f, G)), P(i, G+1) is the position of individual i after the G+1th iteration, β(i, d) is the attraction between individuals i and d, F i G is the scale factor, which follows a normal distribution with a mean of 0.5 and a variance of 0.3; otherwise, the position is updated using the following formula: P(i, G+1) = (1-β(i, b))P(i, G) + β(i, b)(P(b, G) - P(i, G)) + F i G (P(d, G)-P(e, G)); S68: Determine whether the maximum number of iterations has been reached. If so, output the individual position with the smallest objective function value as the result; otherwise, proceed to S63.

2. The method according to claim 1, characterized in that: The first to third CSPs are CSP1_X, and the fourth to eighth CSPs adopt improved CSPs; the improved CSPs include: CBL5 to CBL10, Ghost bottleneck layer, depth convolution, point convolution, average pooling, Sigmoid function, and concat; the input feature maps are respectively input to CBL5 and CBL6; CBL5 is connected to the Ghost bottleneck layer, and the Ghost bottleneck layer outputs two branches, the first branch includes CBL7, depth convolution, and point convolution connected in sequence, and the second branch includes average pooling, CBL8, and Sigmoid function connected in sequence, the output of the Sigmoid function is cross-multiplied with the output of the Ghost bottleneck layer, and the result is input to CBL9, the outputs of CBL6, point convolution, and CBL9 are input to concat, and the output result of concat is input to CBL10.

3. A transmission line insulator fault detection device based on artificial intelligence, characterized in that: It includes data acquisition module, data preprocessing module, model training module and fault detection module, among which: A data acquisition module, used to acquire historical insulator images of a number of transmission lines through a drone, and to mark the status of the insulators in each of the images; A data preprocessing module, used to preprocess the image and use it as a training data set; the preprocessing includes Gaussian blur, Gaussian noise and image sharpening; A model training module, used to build a fault detection model based on YOLO, input the training data set into the model for training, and update the hyperparameters of the model; A fault detection module, used for inputting the image of the transmission line insulator to be detected into the trained fault detection model, and obtaining the corresponding insulator state detection result; The YOLO-based fault detection model includes: a backbone network layer (Backbone layer), a neck layer (Neck layer), and a head layer (Head layer); the model inputs an image into the Backbone layer; the Backbone layer includes a Focus layer, a CFF module, a first CSP, a first CBL, a second CSP, a second CBL, an SPP, and a third CSP connected in sequence; the Neck layer includes a first concat to a fourth concat, a first upsampling, a second upsampling, a third CBL to a tenth CBL, and a fourth CSP to an eighth CSP; the Head layer includes a first feature map P1 to a fourth feature map P4; the output of the first CSP is input to the first concat, and the output of the second CSP is input to the second concat; the fourth CBL, the second upsampling, the second concat, the fourth CSP, the first concat, the second CSP, the third CBL, the fourth CSP ... The third CBL, the first upsampling, the first concat, the fifth CSP, the seventh CBL, and P1 are connected in sequence; the third CBL, the third concat, the seventh CSP, the eighth CBL, and P2 are connected in sequence; the fifth CSP, the fifth CBL, the third concat, the sixth CSP, the sixth CBL, and the fourth concat are connected in sequence; the sixth CSP, the ninth CBL, and P3 are connected in sequence; the third CSP, the fourth CBL, the fourth concat, the eighth CSP, the tenth CBL, and P4 are connected in sequence; the CFF module includes CBL1 to CBL4, a maximum pooling layer and concat; the input feature map is sent to CBL1, and its output is respectively input to CBL2 and the maximum pooling layer, the output of CBL2 is input to CBL3, the output of CBL3 and the maximum pooling layer is input to concat, and the output of concat is input to CBL4; The error function used in the training process is: Where IoU is the cross ratio of the real frame and the predicted frame, d1 is the distance between the center points of the predicted frame and the real frame, c is the minimum diagonal distance between the real frame and the predicted frame, d2 is the distance between the upper left corner coordinates of the predicted frame and the real frame, and r is the ratio of the width of the predicted frame to the real frame; The model uses an improved firefly algorithm to update the hyperparameters of the model and maps the hyperparameters to the positions of fireflies. The algorithm includes: S61: Initialize algorithm parameters, including population size, current iteration number G, maximum iteration number G max , objective function f = ADIOU; S62: Randomly initialize the positions of all individuals in the population; S63: Start iteration: Calculate the Euclidean distance r between individual P(i,G) and individual P(j,G) i,j , P(i,G) is the position parameter of individual i in the Gth iteration; three individuals d, e, and f are randomly selected from the population, denoted as P(d,G), P(e,G), and P(f,G), and they do not overlap with individual i; S64: Calculate the objective function values ​​of all individuals according to the objective function and sort them, and record the individual b with the smallest current objective function value as P(b,G); S65: Calculate the attraction between individuals i and j: where f mean 、f best are the mean and minimum values ​​of the objective function values ​​of all individuals in the current iteration process, β0 is the maximum attraction between individuals, γ is the absorption coefficient, and the scale factor F i G It follows a normal distribution with a mean of 0.5 and a variance of 0.3; S66: Calculate adaptive threshold Δ=1-|f mean -f best | / |f worst -f best |, where f worst It is the maximum value of the objective function of all individuals in the current iteration process; S67: If Δ≥0.5, update the position of the firefly using the following formula: P(i,G+1)=(1-β(i,d))P(i,G)+β(i,d)P(i,G)+F i G (P(e, G)-P(f, G)), P(i, G+1) is the position of individual i after the G+1th iteration, β(i, d) is the attraction between individuals i and d, F i G is the scale factor, which follows a normal distribution with a mean of 0.5 and a variance of 0.3; otherwise, the position is updated using the following formula: P(i, G+1) = (1-β(i, b))P(i, G) + β(i, b)(P(b, G) - P(i, G)) + F i G (P(d, G)-P(e, G)); S68: Determine whether the maximum number of iterations has been reached. If so, output the individual position with the smallest objective function value as the result; otherwise, proceed to S63.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program can be executed by a processor to implement the artificial intelligence-based transmission line insulator fault detection method as described in any one of claims 1-2.

Images