Knowledge reasoning-based high-voltage power line lock bolt defect detection method and system

Abstract

The invention provides a high-voltage power line check bolt defect detection method based on knowledge reasoning, and belongs to the field of bolt defect detection, and the method comprises the following steps: S1, constructing a data set; s2, a defect detection model is constructed, the defect detection model is an improved end-to-end target detection model, and the end-to-end target detection model comprises a model body, an encoder, a decoder and a detection head; s3, constructing a dual-classifier joint reasoning network, wherein the dual-classifier joint reasoning network comprises a lock bolt attribute classifier and a lock bolt defect classifier; and S4, training the check bolt attribute classifier and the check bolt defect classifier, and carrying out defect detection on the bolt by utilizing the trained double classifiers and the reasoning network. Based on the check bolt attribute classifier and the check bolt defect classifier, the position knowledge and the attribute knowledge are combined to assist the bolt defect classifier to detect the bolt defects, and the accuracy of check bolt defect detection is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of anti-loosening bolt defect detection, and in particular relates to a method and system for detecting anti-loosening bolt defects in high-voltage power lines based on knowledge reasoning. Background Art

[0002] As a vital component of the power grid, the safe and stable operation of high-voltage power transmission lines is essential for the smooth flow of electricity. However, due to the large area covered by transmission lines, their constant exposure to various environmental factors, and their diverse composition, timely inspection for abnormalities is one of the most rigorous and demanding tasks in power system operation. Aerial photography technology has rapidly gained popularity in recent years. Using drones to collect images of transmission lines is not only simple to operate but also allows for rapid and secure data collection, saving significant labor costs and improving safety and efficiency. The gradual replacement of manual patrols with drones is an inevitable trend.

[0003] Currently, research using computer vision technology to detect, identify, and classify transmission line defects, such as tower defects, insulator defects, and hardware defects, has achieved promising results. However, relatively little research has been conducted on the detection of defects in high-voltage power line bolts. Bolts are the most common and widely used components on transmission lines. Due to their constant exposure to the external environment, bolts are susceptible to corrosion. Furthermore, since most bolts support and connect transmission lines, they are subject to significant pulling forces, which can eventually cause them to deform or even fall off. Therefore, using computer vision technology to promptly identify whether bolts on transmission lines pose safety risks in power systems has long been a pressing challenge. Summary of the Invention

[0004] In order to address the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a high-voltage power line anti-loosening bolt defect detection method and system based on knowledge reasoning, which uses end-to-end target detection (DETR) as the basic model and expands it using an expansion encoder module to obtain the multi-scale features of the target and infer the position of the anti-loosening bolt through transmission line image relative position encoding (TL-iRPE). Finally, based on the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier, the bolt defect classifier is assisted by combining position knowledge and attribute knowledge to detect bolt defects, thereby further improving the accuracy of bolt defect detection.

[0005] Specifically, the first aspect of the present invention provides a method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning, which includes the following steps:

[0006] S1. Constructing a dataset: Using transmission line images of high-voltage power lines taken by drones, a dataset is constructed and divided into a training set and a test set.

[0007] S2. Build an end-to-end target detection model for defect detection, which includes a dual-classifier joint inference network, an encoder, a decoder, and a detection head. The encoder integrates transmission line location knowledge to construct a relative position code for the transmission line.

[0008] S3. Build a dual-classifier joint inference network, specifically including:

[0009] S31. Construct an anti-loosening bolt attribute classifier: The input of the anti-loosening bolt attribute classifier is the feature vector output by the decoder, and the output is the physical state attributes of the bolt. The physical state attributes of the bolt include gasket integrity, nut missing state, and rust degree.

[0010] S32. Construct an anti-loosening bolt defect classifier: The input of the anti-loosening bolt defect classifier is a concatenated vector of attribute features, position encoding and decoding features, and the output is a bolt defect type, which includes normal, loose, and missing.

[0011] S33, constructing a knowledge constraint matrix to implement logical reasoning, and outputting the bolt defect as the product of the bolt defect type and the knowledge constraint matrix;

[0012] S4. Train the bolt attribute classifier and the bolt defect classifier, and use the dual classifier joint inference network after training to detect defects in anti-loosening bolts.

[0013] Preferably, the anti-loosening bolt property classifier is expressed as:

[0014] S attr =Softmax(W a MLP(F dec ))

[0015] in, is a trainable weight matrix, the output of which corresponds to three physical state attributes, S attr is the physical state property of the bolt, F dec is the decoding feature;

[0016] The anti-loosening bolt defect classifier is expressed as:

[0017]

[0018] in, represents feature splicing, is the defect classification weight, S defect is the bolt defect type, P rel Encode for position;

[0019] The knowledge constraint matrix is expressed as: Output bolt defect S final For: Sfinal =K·S defect .

[0020] Preferably, the physical state attributes of the bolt in step S31 include gasket integrity, nut missing status and rust degree; the bolt defect types in step S32 include normal, loose and missing; in the reasoning process of step S33, when the nut is detected to be missing, set K=0; if the bolt is located at the suspension clamp and the displacement exceeds the preset pixels, the missing judgment weight is increased.

[0021] Preferably, the training process in S4 adopts a multi-task loss function with dynamic weight adjustment:

[0022] L=αL attr +βL defect +γL kl

[0023] Among them, L attr is the attribute classification cross entropy loss, L defect is the defect classification focus loss, L kl is the KL divergence constraint knowledge matrix The consistency with the prior distribution, α, β, γ are hyper parameters.

[0024] Preferably, constructing the encoder in step S2 specifically includes: S21, constructing an encoder module: introducing an expansion encoder module to improve the adaptability of the model to hardware and bolts of different sizes;

[0025] S22. Integrate the transmission line location knowledge to construct the transmission line relative position coding. The transmission line relative position coding describes the topological relationship between bolts through the dual dimensions of distance and angle.

[0026] Preferably, in step S21, the encoder module includes a projection layer and an expanded convolution residual block. The projection layer uses a 1×1 convolution layer for channel dimensionality reduction, and a 3×3 convolution layer is superimposed to refine the contextual semantic information. The expanded convolution residual block generates a multi-level receptive field output feature containing all target scales by superimposing multiple continuous expanded convolution residual blocks with different expansion rates. The corresponding features of the original scale range and the expanded scale range are added and fused through the residual connection, and finally the feature containing the multi-level receptive field is output.

[0027] Preferably, the relative position encoding calculation in step S22 is as follows: S221, using Cartesian coordinate transformation, for any two elements in the image, calculate the relative displacement; convert the relative displacement into the distance r in the polar coordinate system ij and angle θ ij :S222, introduce angle-related weight matrix Map the angle information into a feature vector; S223, discretize the continuous distance using a piecewise index function; S224, generate relative position encoding and embed it into the self-attention mechanism; S225, fuse the encoding result with the self-attention weight matrix through addition.

[0028] Preferably, in step S223, the continuous distance is discretized using a piecewise index function as follows:

[0029] g(r ij )=log(r ij +ò)·s; where ò is the smoothing factor and s is the scale factor;

[0030] In step S224, the relative position encoding is generated and embedded in the self-attention mechanism using the following formula:

[0031] P rel =MLP(r ij +Wθ·θ ij ); where r ij is the embedding vector of the distance bucket index, and MLP is a multi-layer perceptron.

[0032] Preferably, in S225, the encoding result is fused with the self-attention weight matrix by addition as follows:

[0033]

[0034] In a second aspect, the present invention provides a bolt defect detection system for a high-voltage power line anti-loosening bolt defect detection method based on knowledge reasoning, which includes a data set construction unit, a defect detection model construction unit, a dual-classifier joint inference network construction unit, a training unit, and a defect detection unit;

[0035] The data set construction unit is used to construct a data set using the transmission line images taken by the UAV, and divide the data set into a training set and a test set;

[0036] The defect detection model construction unit is used to construct a defect detection model. The defect detection model is an improved end-to-end target detection model. The end-to-end target detection model includes a dual-classifier joint inference network, an encoder, a decoder, and a detection head.

[0037] The dual-classifier joint reasoning network construction unit is used to construct a dual-classifier joint reasoning network, which includes an anti-loosening bolt attribute classifier and an anti-loosening bolt defect classifier;

[0038] The training unit is used to train the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier;

[0039] The defect detection unit uses the trained dual classifier joint inference network to detect defects on bolts.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) This paper proposes an end-to-end anti-loosening bolt defect detection method based on knowledge reasoning. This end-to-end defect detection model reduces inference time by 50%, supports real-time detection, and effectively improves detection efficiency. Due to the integration of the transmission line knowledge reasoning mechanism, the detection accuracy of visually indistinguishable defects is effectively improved. The expansion encoder module effectively captures bolt targets of different sizes, enhances the model's adaptability at multiple scales, reduces the missed detection rate, and significantly improves the detection rate of bolt faults.

[0042] (2) The present invention proposes to improve the end-to-end model by extending the encoder. In order to solve the problem that the traditional end-to-end target detection model (DETR) performs poorly in small target detection, the extended encoder module is introduced to improve the accuracy of small target detection. At the same time, the relative position coding of the transmission line is constructed. The relative position coding of the transmission line describes the topological relationship between the anti-loosening bolts through the dual dimensions of distance and angle. The relative position coding of the transmission line can further improve the accuracy of the model.

[0043] (3) The present invention constructs an anti-loosening bolt attribute classifier and an anti-loosening bolt defect classifier. Based on the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier, the anti-loosening bolt defect classifier is assisted in detecting bolt defects by combining position knowledge and attribute knowledge, thereby further improving the accuracy of bolt defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the flow of the high-voltage power line anti-loosening bolt defect detection method based on knowledge reasoning of the present invention;

[0045] Figure 2 It is a schematic diagram of the workflow of the present invention;

[0046] Figure 3 This is a block diagram of a bolt defect detection system for a high-voltage power line anti-loosening bolt defect detection method based on knowledge reasoning according to the present invention. DETAILED DESCRIPTION

[0047] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0048] In the first aspect, the present invention provides a method for detecting defects in anti-loosening bolts of high-voltage power lines based on knowledge reasoning, such as Figure 1 and Figure 2 As shown, it includes the following steps:

[0049] S1. Constructing a Dataset: A dataset is constructed using transmission line images captured by drones and divided into a training set and a test set. In this embodiment, drones are used to capture and annotate high-voltage power line images to obtain various types of bolt defects. The dataset is then divided into a training set and a test set based on the proportions.

[0050] S2. Construct a defect detection model. The defect detection model is an improved end-to-end target detection model. The end-to-end target detection model includes a backbone, an encoder, a decoder, and a detection head. The backbone is a dual-classifier joint inference network. The dual-classifier joint inference network extracts input image features and extends the image features. The self-attention mechanism in the encoder is responsible for learning image features and then transferring the optimized features to the decoder. The decoder generates N object query vectors through initialization. The object query vectors are first optimized through the self-attention mechanism of the decoder, then fused with the image features output by the encoder, and then optimized again through the cross-self-attention mechanism. Finally, the output of the decoder is the optimized object query vector. In the final stage, the detection head of DETR uses a bipartite graph matching method to apply these object query vectors to the classification task and bounding box regression task of target detection respectively.

[0051] This step specifically includes the following sub-steps:

[0052] S21. Construct encoder module: Introduce an expansion encoder module to improve the model's adaptability to hardware and bolts of different sizes. The receptive field calculation formula of the expansion encoder module is:

[0053] S=(rate-1)×(s-1)+s

[0054] Here, S represents the size of the dilated convolution receptive field, rate is the dilation rate (which controls the convolution kernel sampling interval), and s represents the convolution kernel size. Hardware generally refers to metal fixtures used on power lines, and bolts connect hardware and fasten conductors.

[0055] The receptive field (RF) refers to the input area that a neuron in a neural network can "see" or respond to. Specifically, the larger the RF of a neuron, the richer and more global the image information it can capture. This helps the model understand contextual information within a wider range, thereby improving its ability to recognize objects.

[0056] In object detection tasks, expanding the receptive field can help the model better capture objects of varying scales. For example, by using dilated convolutions, the receptive field of the feature map can be expanded without increasing computational cost, allowing the model to simultaneously cover a wider range of object scales. This is particularly useful for detecting objects of small size or with large positional variations. Therefore, properly designing the network structure to adjust the receptive field size is one of the key factors in improving model performance.

[0057] The encoder module consists of two parts: a projection layer and a residual block. The projection layer uses a 1×1 convolution layer to perform channel dimensionality reduction, and then a 3×3 convolution layer is superimposed to refine the contextual semantic information. Although the dilation operation expands the receptive field by multiplying the original coverage scale by a coefficient greater than 1, it still cannot cover all target scales. For this reason, a dilated convolution residual block is specially designed for the dilated encoder module. This residual block generates multi-receptive field output features covering all target scales by superimposing four consecutive dilated convolution residual blocks with different dilation rates. Finally, the corresponding features of the original scale range and the expanded scale range are added and fused through the residual connection, and the final output contains features of multi-level receptive fields. This model only needs to use the C5 layer features to cover all target scales, while obtaining feature maps with richer semantic information and significantly reducing the amount of computation. The module convolution kernel dilation rate is set to [1,2,4,8], and the mathematical expression is:

[0058] F multi =DConv(F base ; rate = r)

[0059] Among them, F base It is the basic feature output by ResNet50, and DConv represents the dilated convolution operation.

[0060] S22. Integrate transmission line location knowledge to construct a relative position code for transmission lines. This code describes the topological relationship between bolts using both distance and angle. This can further improve model accuracy.

[0061] Preferably, the specific parameters in step S22 are calculated as follows:

[0062] S221, using Cartesian coordinate transformation, for any two elements i(x i ,y i ) and j(x j ,y j ), calculate the relative displacement:

[0063] Δx=x i -x j ,Δy=y i -y j

[0064] At the same time, the relative displacement is converted into the distance r in the polar coordinate system ij and angle θ ij :

[0065]

[0066] S222, introduce angle-related weight matrix Map angle information into feature vectors.

[0067] S223. Use piecewise index function to discretize continuous distance:

[0068] g(r ij )=log(r ij +ò)·s

[0069] Where, ò is the smoothing factor and s is the scale factor. In specific applications, the smoothing factor ò is generally 10-5.

[0070] S224, relative position encoding is generated and embedded in the self-attention mechanism through the following formula:

[0071] P rel =MLP(r ij +W θ ·θ ij )

[0072] Among them, r ij is the embedding vector of the distance bucket index, and MLP is a multi-layer perceptron.

[0073] S225. Fuse the encoding result with the self-attention weight matrix by addition:

[0074]

[0075] S3. Construct a dual-classifier joint inference network, which includes a bolt property classifier and a bolt defect classifier. This includes the following sub-steps:

[0076] S31. Construct an anti-loosening bolt attribute classifier: The input of the bolt attribute classifier is the feature vector output by the decoder. The output is the physical state attributes of the bolt, which include the integrity of the gasket, the missing state of the nut, and the degree of rust. The anti-loosening bolt attribute classifier is expressed as:

[0077] S attr =Softmax(W a MLP(F dec ))

[0078] in, is a trainable weight matrix, the output dimension of the trainable weight matrix corresponds to three physical state attributes, S attr is the physical state property of the bolt, F dec Decoding features.

[0079] S32. Construct an anti-loosening bolt defect classifier: The input of the anti-loosening bolt defect classifier is the attribute feature S attr , position code P rel and decoding feature F dec The splicing vector of is output as the bolt defect type, which includes normal, loose and missing. The bolt defect classifier model is expressed as:

[0080]

[0081] in, represents the feature concatenation operation, is the defect classification weight, S defect The defect type of bolt.

[0082] S33. Constructing a knowledge constraint matrix Implement logical reasoning and output bolt defects:

[0083] S final =K·S defect

[0084] Among them, S final is the output bolt defect, and K is the knowledge constraint matrix. During the inference process, when a missing nut is detected, K is set to 0. If the bolt is located at the suspension clamp and the displacement exceeds a preset number of pixels, the missing judgment weight is increased. When a missing nut is detected, that is, the attribute classifier outputs the "missing nut" status, the knowledge constraint matrix K is set to 0, and the matrix multiplication output is 0. Increasing the missing judgment weight means adjusting the classifier's sensitivity to the "missing" category. Specific implementation methods include:

[0085] Position and displacement condition trigger: When the bolt is located at the suspension clamp, that is, it is subjected to large mechanical stress, and the displacement exceeds the set pixel threshold, the system considers the loosening risk to be extremely high.

[0086] Dynamic weight adjustment mechanism: During the training phase, the loss weight of the "missing" category is increased through FocalLoss, so that the model pays more attention to difficult samples (such as slightly loose bolts); during the inference phase, the output of the defect classifier is corrected through the weighting of the knowledge constraint matrix.

[0087] S4, train the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier; where L attr is the attribute classification cross entropy loss, L defect is the defect classification focus loss, Lkl is the KL divergence constraint knowledge matrix The consistency with the prior distribution, α, β, and γ are hyperparameters. Finally, the trained dual-classifier joint inference network is used to detect defects in bolt images collected by the drone in real time, outputting defective bolts and the bolt defect type.

[0088] In the second aspect, the present invention provides a bolt defect detection system for a high-voltage power line anti-loosening bolt defect detection method based on knowledge reasoning, such as Figure 3 As shown, it includes a data set construction unit 1, a defect detection model construction unit 2, a dual classifier joint inference network construction unit 3, a training unit 4 and a defect detection unit 5.

[0089] The data set construction unit 1 is used to construct a data set using the transmission line images taken by the drone, and divide the data set into a training set and a test set.

[0090] The defect detection model construction unit 2 is used to construct a defect detection model. The defect detection model is an improved end-to-end target detection model. The end-to-end target detection model includes a dual-classifier joint inference network, an encoder, a decoder and a detection head.

[0091] The dual-classifier joint reasoning network construction unit 3 is used to construct a dual-classifier joint reasoning network, which includes an anti-loosening bolt attribute classifier and an anti-loosening bolt defect classifier.

[0092] The training unit 4 is used to train the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier.

[0093] The defect detection unit 5 uses the trained dual classifier joint inference network to perform defect detection on the bolts.

[0094] This embodiment was tested on the VIBD dataset (a visually indistinguishable bolt defect detection dataset) to verify the object detection performance and defect detection performance of PA-DETR. Experimental results show that in the VIBD dataset, this model can effectively identify different types of accessories, bolts and their defects. Test results on the MS-COCO 2017 dataset (a standard dataset for object detection tasks) show that the proposed method improves the AP from 42.0% to 46.5% compared to the popular two-stage model Faster-RCNN, and improves the AP value by 0.1% compared to the most advanced single-stage model YOLOX-M. Compared with the basic model DETR-ResNet50, the model with the expansion encoder added has an AP increase of 4.5%. In multiple detection scenarios, the missed detection rate is reduced, and the detection rate of bolt faults is greatly improved.

[0095] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning, characterized by: It includes: S1. Constructing a dataset: Using transmission line images of high-voltage power lines taken by drones, a dataset is constructed and divided into a training set and a test set. S2. Build an end-to-end target detection model for defect detection, which includes a dual-classifier joint inference network, an encoder, a decoder, and a detection head. The encoder integrates transmission line location knowledge to construct a relative position code for the transmission line. S3. Build a dual-classifier joint inference network, specifically including: S31. Construct an anti-loosening bolt attribute classifier: The input of the anti-loosening bolt attribute classifier is the feature vector output by the decoder, and the output is the physical state attributes of the bolt. The physical state attributes of the bolt include gasket integrity, nut missing state, and rust degree. S32. Construct an anti-loosening bolt defect classifier: The input of the anti-loosening bolt defect classifier is a concatenated vector of attribute features, position encoding and decoding features, and the output is a bolt defect type, which includes normal, loose, and missing. S33, constructing a knowledge constraint matrix to implement logical reasoning, and outputting the bolt defect as the product of the bolt defect type and the knowledge constraint matrix; S4. Train the bolt attribute classifier and the bolt defect classifier, and use the dual classifier joint inference network after training to detect defects in anti-loosening bolts.

2. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 1 is characterized in that: The anti-loosening bolt attribute classifier is expressed as: S attr =Softmax(W a MLP(F dec )); in, is the trainable weight matrix, S attr is the physical state property of the bolt, F dec is the decoding feature; the anti-loosening bolt defect classifier is expressed as: in, represents feature splicing, is the defect classification weight, S defect is the bolt defect type, P rel is the position encoding; the knowledge constraint matrix is expressed as: Output bolt defect S final For: S final =K·S defect .

3. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 2 is characterized in that: During the reasoning process of step S33, when it is detected that the nut is missing, K=0; if the bolt is located at the suspension clamp and the displacement exceeds the preset pixels, the missing judgment weight is increased.

4. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 1 is characterized in that: The training process in S4 uses a multi-task loss function with dynamic weight adjustment: L=αL attr +βL defect +γL kl Among them, L attr is the attribute classification cross entropy loss, L defect is the defect classification focus loss, L kl is the KL divergence constraint knowledge matrix The consistency with the prior distribution, α, β, γ are hyper parameters.

5. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 1 is characterized in that: Constructing the encoder in step S2 specifically includes: S21, constructing an encoder module: introducing an expansion encoder module to improve the adaptability of the model to hardware and bolts of different sizes; S22. Integrate the transmission line location knowledge to construct the transmission line relative position coding. The transmission line relative position coding describes the topological relationship between bolts through the dual dimensions of distance and angle.

6. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 4 is characterized in that: In step S21, the encoder module includes a projection layer and an expanded convolution residual block. The projection layer uses a 1×1 convolution layer for channel dimensionality reduction and superimposes a 3×3 convolution layer to refine the contextual semantic information. The expanded convolution residual block generates a multi-level receptive field output feature containing all target scales by superimposing multiple continuous expanded convolution residual blocks with different expansion rates. The corresponding features of the original scale range and the expanded scale range are added and fused through the residual connection, and finally the features containing the multi-level receptive field are output.

7. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 4 is characterized in that: The relative position code calculation in step S22 is as follows: S221, using Cartesian coordinate transformation, for any two elements in the image, calculate the relative displacement; convert the relative displacement into the distance r in the polar coordinate system ij and angle θ ij :S222, introduce angle-related weight matrix Map the angle information into a feature vector; S223, discretize the continuous distance using a piecewise index function; S224, generate relative position encoding and embed it into the self-attention mechanism; S225, fuse the encoding result with the self-attention weight matrix through addition.

8. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 4 is characterized in that: In step S223, the continuous distance is discretized using the piecewise index function as follows: in, is the smoothing factor, s is the scale factor; In step S224, the relative position encoding is generated and embedded in the self-attention mechanism using the following formula: P rel =MLP(r ij +Wθ·θ ij ); where r ij is the embedding vector of the distance bucket index, and MLP is a multi-layer perceptron.

9. The method for detecting defects in high-voltage power line anti-loosening bolts based on knowledge reasoning according to claim 5 is characterized in that: In S225, the encoding result is fused with the self-attention weight matrix by addition as follows:

10. A bolt defect detection system for the high-voltage power line anti-loosening bolt defect detection method based on knowledge reasoning according to any one of claims 1 to 9, characterized in that: It includes a data set construction unit, a defect detection model construction unit, a dual classifier joint reasoning network construction unit, a training unit and a defect detection unit; The data set construction unit is used to construct a data set using the transmission line images taken by the UAV, and divide the data set into a training set and a test set; The defect detection model construction unit is used to construct a defect detection model. The defect detection model is an improved end-to-end target detection model. The end-to-end target detection model includes a dual-classifier joint inference network, an encoder, a decoder, and a detection head. The dual-classifier joint reasoning network construction unit is used to construct a dual-classifier joint reasoning network, which includes an anti-loosening bolt attribute classifier and an anti-loosening bolt defect classifier; The training unit is used to train the anti-loosening bolt attribute classifier and the anti-loosening bolt defect classifier; The defect detection unit uses the trained dual classifier joint inference network to detect defects on bolts.

Images