CN117095311A - Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal - Google Patents

Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal Download PDF

Info

Publication number
CN117095311A
CN117095311A CN202310028596.5A CN202310028596A CN117095311A CN 117095311 A CN117095311 A CN 117095311A CN 202310028596 A CN202310028596 A CN 202310028596A CN 117095311 A CN117095311 A CN 117095311A
Authority
CN
China
Prior art keywords
network
hot spot
feature
information
global
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310028596.5A
Other languages
Chinese (zh)
Inventor
郝帅
何田
马旭
李嘉豪
王海莹
吴瑛琦
赵秋林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Science and Technology
Original Assignee
Xian University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian University of Science and Technology filed Critical Xian University of Science and Technology
Priority to CN202310028596.5A priority Critical patent/CN117095311A/en
Publication of CN117095311A publication Critical patent/CN117095311A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/17—Terrestrial scenes taken from planes or by drones
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Remote Sensing (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention belongs to the field of intelligent photovoltaic inspection, and discloses an intelligent photovoltaic hot spot fault detection method, an intelligent photovoltaic hot spot fault detection system, an intelligent photovoltaic hot spot fault detection medium, an intelligent photovoltaic hot spot fault detection equipment and an intelligent photovoltaic hot spot fault detection terminal, wherein the intelligent photovoltaic hot spot fault detection method comprises the following steps: the knowledge distillation module for cooperative training of teachers and students is constructed, so that the detection accuracy is ensured, and the algorithm reasoning efficiency is improved; a backbone network based on CSPHB is designed to enhance the characteristic expression capability of the target to be detected; a BiMAP model is built, and input features are fused from the global and local angles by adopting a parallel method, so that the aggregation capability of target features is enhanced; meanwhile, a CgT module is provided, so that the detection system can still find out hot spot faults in time and remove and overhaul the hot spot faults under various severe environments. According to the invention, the unmanned aerial vehicle inspection technology, the sensor technology and the target detection algorithm are subjected to cross fusion, so that the large-scale, rapid, fine and intelligent operation and maintenance tasks of the photovoltaic power station are automatically completed, the power generation efficiency of the photovoltaic power station is improved, and the operation and maintenance cost of the photovoltaic power station is reduced.

Description

Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal
Technical Field
The invention belongs to the technical field of intelligent photovoltaic inspection, and particularly relates to an intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal.
Background
At present, with the rapid development of economy and society, the phenomenon of energy shortage in China is increasingly severe, and the traditional fossil energy sources face a plurality of problems such as non-renewable resources, serious pollution and the like. For this reason, the supply pattern of energy in China is changing from fossil energy to renewable clean energy, and the new clean energy industry is greatly developed. Among them, solar photovoltaic power generation is a key technology in the new energy field, and has the advantages of green low carbon, flexible use mode, wide distribution area, and the like, and has been paid attention to widely.
The photovoltaic power generation system can directly convert solar radiation energy into electric energy by utilizing the photovoltaic effect of the solar cell semiconductor material so as to be used by users. However, in the long-term use process, the photovoltaic module is inevitably covered by dust, fallen leaves and other shielding materials, so that local shadow phenomenon is caused, the current and the voltage of certain battery singlechips are changed, and further the local power of the photovoltaic module is increased, and a hot spot effect is generated. The hot spot effect has a local temperature rise characteristic, and when severe, the hot spot effect fuses welding spots and damages grid lines, so that the service life of the photovoltaic module is shortened. Meanwhile, the hot spot effect enables part of normal battery boards to be shielded and not work normally, so that the production efficiency of the large photovoltaic power station is reduced. Therefore, in order to ensure that the photovoltaic power generation system can stably operate for a long period of time, it is necessary to periodically perform hot spot failure detection of the photovoltaic module.
Under the support of increasingly mature electric power technology, photovoltaic power stations are widely distributed in complex zones with wide regions and sufficient sunlight, such as mountain power stations, water power stations and the like. The coverage area is huge, and the coverage area is generally in a messy and scattered shape due to the limitation of the terrain, so that great challenges are brought to the maintenance work of the power system. The traditional manual inspection mode is limited by operation and maintenance cost, working conditions and labor efficiency, and is difficult to meet the operation and maintenance work of the photovoltaic module with high precision and high efficiency. With the development of unmanned aerial vehicle industry, the development of computer vision technology and the increasingly mature of intelligent photovoltaic inspection system, through carrying on sensors such as visible light, thermal infrared and carry out photovoltaic module hot spot fault detection, can accomplish photovoltaic power plant on a large scale, fast, meticulously, intelligent operation maintenance task automatically. Compared with the traditional manual inspection work mode, the unmanned aerial vehicle inspection mode can effectively improve the accuracy rate of hot spot detection and inspection efficiency.
The photovoltaic hot spot fault detection algorithm is used as a key technology of the intelligent photovoltaic inspection system and can be mainly divided into two major categories, namely a fault diagnosis method based on electrical characteristics and an image diagnosis method based on a computer vision technology. The fault diagnosis method based on the electrical characteristics monitors and analyzes electrical data such as output voltage, output current and output power of the photovoltaic module by using a mathematical model method or machine learning, and can accurately diagnose hot spot faults in the photovoltaic module. However, the algorithm needs to be added with auxiliary detection of peripheral circuits, has higher cost and lower efficiency, and is difficult to meet the operation and maintenance requirements of actual large-scale photovoltaic power stations.
Image diagnosis methods using computer vision techniques as a core can be classified into two types of conventional image processing algorithms and deep learning algorithms. The traditional hot spot fault detection algorithm utilizes a sliding window technology to realize manual feature extraction and combines a classification identifier to complete a fault detection task, and the algorithm can accurately detect a target in certain specific scenes, but is difficult to capture advanced semantic information of the hot spot target, and has poor generalization capability in a complex environment. Compared with the traditional hot spot fault detection algorithm based on artificial feature extraction, the hot spot image diagnosis algorithm based on deep learning can automatically learn target features by utilizing excellent feature extraction and nonlinear fitting capability of a convolutional neural network, so that whether a hot spot target exists in an image or a video sequence or not is accurately judged, accurate positioning is achieved, and excellent detection effects are achieved in terms of detection precision, speed and generalization capability. However, in the existing stage, various photovoltaic hot spot fault detection methods generally consider that algorithm detection precision and model weight reduction cannot be considered simultaneously, so that in order to ensure inspection performance, a large detection network with high calculation cost is generally selected to perform fault positioning and identification on photovoltaic inspection data. Although the accuracy of photovoltaic power station operation maintenance is improved by the mode, the inspection efficiency is generally low, and the real-time detection requirement cannot be met. Meanwhile, due to the fact that the shooting angle, the height and the performance of the unmanned aerial vehicle are limited, the collected photovoltaic data are fuzzy, and most of hot spot faults are small-scale feature distortion targets, the feature of the target to be detected is difficult to effectively extract by a fault detection algorithm, the detection performance of infrared small-scale hot spot faults is limited, and further network reliability cannot be guaranteed.
Therefore, how to consider the real-time efficiency and algorithm precision of the detection network, improve the characteristic expression capability of the infrared hot spot target, improve the small-scale target detection performance and reduce the network false alarm rate and false alarm rate is a technical problem to be solved in the current stage.
Through the above analysis, the problems and defects existing in the prior art are as follows:
(1) Photovoltaic power plants have been widely distributed in complex regions of wide territories and abundant sunlight, such as mountain power stations, water power stations, and the like. The coverage area is huge, and the coverage area is generally in a messy and scattered shape due to the limitation of the terrain, so that great challenges are brought to the maintenance work of the power system.
(2) The traditional manual inspection mode is limited by operation and maintenance cost, working conditions and labor efficiency, and is difficult to meet the operation and maintenance tasks of the photovoltaic module with high precision and high efficiency.
(3) The fault diagnosis method based on the electrical characteristics needs to increase auxiliary detection of peripheral circuits, has higher cost and lower efficiency, and is difficult to meet the operation and maintenance requirements of the actual large-scale photovoltaic power station.
(4) The traditional hot spot image detection algorithm is difficult to capture the high-level semantic information of the hot spot target, so that the target can be accurately detected only in certain specific scenes, and the generalization capability is poor in a complex environment.
(5) The algorithm detection performance is generally in positive correlation with the detection model parameter and the network depth, but the larger the model parameter is, the deeper the network depth is, the larger the algorithm calculation cost is correspondingly, and the longer the running time is. Therefore, the combination of detection precision and model weight reduction is a great challenge of the current intelligent photovoltaic inspection algorithm.
(6) Because unmanned aerial vehicle shooting angle is changeable, easily cause the shooting image to appear characteristic distortion, noise serious scheduling problem to lead to hot spot target feature to be difficult to effectively express, and then influence final detection precision.
(7) The unmanned aerial vehicle hot spot target is affected by objective factors such as shooting height of the unmanned aerial vehicle, geographical position where data are located and the like, the hot spot target is small in general size, is disordered in distribution, and lacks appearance information such as textures, shapes and colors, so that the prior anchor point frame is weak in adaptability to the small-size target, and the problems of false alarm and false alarm occur.
(8) Under a dense multi-target scene, all targets appearing in images or videos are difficult to accurately detect by various intelligent photovoltaic inspection algorithms at the present stage, and the problem of missed detection and false detection at different degrees exists, so that the reliability of an inspection system is lower, and the intelligent photovoltaic inspection performance is directly limited.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides an intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal.
The invention is realized in such a way that an intelligent photovoltaic hot spot fault detection method comprises the following steps:
from the viewpoint of model weight reduction, a knowledge distillation module for cooperative training of teachers and students is constructed, so that the model reasoning efficiency is ensured, and the algorithm detection precision is improved; in consideration of the positive correlation between target feature extraction and detection performance, a backbone network based on CSPHN is designed, and the feature representation capability of an algorithm is remarkably enhanced under the condition of not consuming excessive network memory by combining the respective advantages of CSPNet and HorNet structures; inspired by the principle of visual nerve layering cognition, a BiMAF model is constructed, and the aggregation capability of a neck network on target features is enhanced by fusing input features from the global and local angles by adopting a parallel method; meanwhile, in order to further enhance the detection performance of the network on the dense small targets, a CgT module is provided before decoupling the prediction head, and an attention matrix is generated by mining static and dynamic context information, so that the detection system can still timely discover hot spot faults existing in the photovoltaic power generation system in a dense small target scene and remove and overhaul the hot spot faults, thereby ensuring safe and stable operation of the photovoltaic power generation system.
Further, the intelligent photovoltaic hot spot fault detection method comprises the following steps:
step one, designing a new mode of cooperative training of a teacher and a student network based on a knowledge distillation idea. The method has the advantages that the detection precision of the algorithm is improved by means of the advantages of the teacher depth detection network, the small parameter quantity characteristics of the student network are combined, the reasoning efficiency of the algorithm is improved, and the detection precision and the light weight requirement of the model are further met;
secondly, building a student network based on a YOLOX algorithm aiming at an unmanned aerial vehicle intelligent patrol data set provided by a photovoltaic power station operation and maintenance company, greatly reducing the number of network parameters by building an anchor-free frame model, and improving the detection precision of the algorithm and simultaneously accelerating the network convergence speed by combining a decoupling prediction head;
thirdly, considering the positive correlation between target feature extraction and detection performance, constructing a CSPHN (HorNet-based cross stage partial network) module in a teacher backbone network, and obtaining high-order space interaction information similar to a Transformer by utilizing gated convolution and recursive design on the basis of keeping beneficial inductive bias of a convolutional neural network so as to enhance the feature expression capability of an algorithm on a hot spot target;
and fourthly, in order to acquire more discriminative target features from the high-level semantic information and the low-level positioning information output by the backbone network, a BiMAF (Bi-branch multi-level feature adaptive fusion) module is designed in the neck network of the teacher. The multi-level characteristics are aggregated from two angles of global and local in a parallel fusion mode, so that the network can be selectively focused on the area containing the target significance information, and the characteristic aggregation capability of the neck network is further enhanced;
Fifth, to further improve the detection performance of the small-scale target, cgT (g) is constructed before the teacher decouples the prediction head n conv-based contextual Transformer) module, by executing self-attention learning operation and mining context information among two-dimensional feature map input keys, the expression capability of a transducer architecture is improved, and further, the hot spot small-scale target detection precision under various dense scenes is improved.
Further, the construction of the knowledge distillation co-training new mode comprises the following steps:
(1.1) aiming at the unbalanced problem of foreground and background characteristics of the teacher and student networks, designing a local distillation function to separate the foreground and background information of the images and guide the student networks to pay attention to important pixels and channel characteristics;
(1.2) local distillation cuts off the correlation between the image front and back backgrounds, which makes it difficult for the detection network to capture the image global information, limiting network detection performance. Therefore, a global distillation function is proposed to reconstruct the relation between different pixels and transfer the relation from a teacher network back to a student network, so as to compensate the global information lost in the local distillation process;
and (1.3) combining two modes of local distillation and global distillation, thereby inheriting information in a large teacher network into a compact student network, and ensuring that the information does not add extra cost in the reasoning process to obtain strong performance.
Further, the specific process of the step (1.1) comprises the following steps:
(1.1.1) construction of a feature map F for i, j in the abscissa and ordinate, respectivelyBinary mask M i,j Separating the foreground and background of the image, as shown in the following formula:
wherein r is a target real frame;
(1.1.2) setting a scale mask S for balancing the loss of the detection network to different scale targets and foreground areas i,j ：
Wherein H is r W and W r Respectively representing the height and the width of a target real frame;
(1.1.3) construction of a spatial attention maskChannel attention mask->To improve model distillation performance:
wherein H, W and C represent feature height, width and channel, respectively; t is a temperature super parameter; f (F) c And F i,j Characteristic information of a c-th channel and characteristic information of a size of i×j are respectively represented;
(1.1.4) during training, binary mask M is used i,j Scale mask S i,j And attention mask A S (F) A is a C (F) Guiding students to learn teacher network key space and channel information through network, thereby constructing the following characteristic loss function L fea Attention loss function L at ：
In the method, in the process of the invention,and->Characteristic diagrams respectively representing teacher and student networks; f (·) is to F S Adjust to F T Reconstruction operators of the same dimension; alpha, beta and gamma are super parameters for balancing various losses; l represents l 1 A norm operator; />And-> Respectively representing a spatial attention mask and a channel attention mask corresponding to the teacher and the student network;
(1.1.5) obtaining a final local distillation function by calculating the characteristic loss and the attention loss:
L focal ＝L fea +L at
further, the specific process of the step (1.2) comprises:
capturing global information of an image by using a GcBLock module so as to obtain global loss L global ：
L global ＝λ·∑(R(F T )-R(F S )) 2
Wherein λ represents a balance loss hyper-parameter; r (F) represents feature global information, which can be expressed as:
in which W is k ，W v1 And W is v2 Representing a convolution layer; LN representation layer normalization; n (N) p Representing the feature pixel count; reLU represents a linear activation function.
Further, the design of the student network includes:
the YOLOX network includes a Backbone network (Backbone), a Neck network (neg) and decoupling pre-measurement heads;
(1) The Backbone is mainly composed of three parts, focus, CSPNet (Cross Stage Partial Network) and SPP (Spatial Pyramid Pooling). The Focus slicing module not only enlarges the network receptive field, but also can effectively inhibit the loss of image characteristic information, thereby accelerating the training speed. The CSPNet structure aims to solve the problem that the calculation cost is too high due to repeated gradient information in the network optimization process. The SPP structure expands the backbone network receptive field by using multistage pooling operation, and improves the multi-scale feature fusion capability.
(2) The Neck network builds a CSP2_X structure by referring to the CSPNet to strengthen the network feature fusion capability, and utilizes the FPN+PAN feature pyramid structure to conduct feature aggregation on different scale information. Wherein, FPN transmits strong semantic information from top to bottom and PAN transmits strong positioning information from bottom to top.
(3) And finally, constructing a decoupling pre-measuring head, and decoupling the classifying branch of the focusing texture information and the positioning branch of the focusing edge information, so that the problem of space dislocation is effectively solved, and the convergence speed is improved.
Further, the constructing of the CSPHN module includes:
(1) From the network structure design perspective, the feature mapping of the input layer is divided into two parts, and the two parts are combined by utilizing a cross-stage hierarchical structure, so that the network memory occupation amount is reduced, and the learning capacity of the convolutional neural network is enhanced.
(2) In order to make the backbone network have both local and global feature extraction capability, the residual unit in the original CSPNet model is replaced by a HorNet module. The module inherits elements of a transducer model space mixing layer and a feedforward network cascadeArchitecture and using gated recursive convolution (g n Conv) captures the high-order space interaction in the feature map, so that the secondary calculation complexity caused by multiple dot products in the execution process of the multi-head attention mechanism can be avoided, and the extraction capability of the backbone network to the detection target features is improved.
The specific process comprises the following steps:
first, let g n Conv input features areThen a set of projection features p 0 Is->Can be obtained by the following formula:
wherein phi (·) represents the projection operator, an
Then, a gated recursive convolution is performed, the formula:
p k+1 ＝DW k (q k )⊙g k (p k )/α,k＝0,1,...,n-1
where α is a scaling factor, { DW k The } represents a set of deep convolutions, ++represents a dot-product operation, and { g } is k The dimensions used to match the unused sequences are shown as follows:
finally, performing feature projection after the top-level recursion operation to obtain g n Output result of Conv.
Further, the design of the BiMAF model includes:
(2.1) global feature fusion tributaries: global pooling is typically used to globally encode spatial information, but it compresses global spatial informationIs condensed into the channel descriptor, making it difficult to retain feature location information. Thus, to enable the feature fusion module to capture the remote spatial interaction features with precise location information, the global pooling operation is converted into a pair of dimensional feature operators. Then, in order to fully utilize the feature information obtained after the coordinate information embedding operation, the feature information is subjected to cascading and convolution operation, and global feature weights are output by using a sigmoid normalization function. Finally, obtaining a global feature fusion result Output through weight distribution operation g 。
(2.2) constructing local feature weights by using a 1×1 convolution and sigmoid normalization function, and obtaining a local feature fusion result Output by using a weight distribution operation l 。
And (2.3) aggregating global and local feature output by the BiMAF model in a parallel fusion mode to obtain a final feature fusion result.
Output＝[Output g ,Output l ]
Further, the calculating process of the global feature fusion result in the step (2.1) includes:
(2.1.1) assume that the multi-level feature inputs are respectivelyThe elements and operations are performed as follows:
M＝Input 1 +Input 2
encoding each channel in the horizontal and vertical directions by using two pooling cores with the sizes of (H, 1) and (1, W) to obtain the output of the c-th channel with the height of HAnd the output of the c-th channel with width w +.>
(2.1.2) performing the following concatenation and convolution operations:
f＝δ(F 1 ([z h ,z w ]))
in [ z ] h ,z w ]Representing a join operation along a spatial dimension; delta is the activation function; f is an intermediate feature map obtained by encoding in the horizontal and vertical directions; f represents a 1×1 convolution operation;
by decomposing f along the spatial dimension, two independent tensors are obtained: f (f) h ∈R C/r×H F h ∈R C/r×H Wherein r is the downsampling ratio; at the same time, two 1X 1 convolutions F are performed h And F w Operate to ensure that two independent tensors have the same number of channels, thereby obtaining global feature weights W g ：
W g ＝sigmoid(F h ([f h ]))×sigmoid([f w ])
(2.1.3) Output of Global feature fusion branch g The method comprises the following steps:
Output g ＝Input 1 ×W g +Input 2 ×(1-W g )
the calculation process of the local feature fusion result in the step (2.2) comprises the following steps:
carrying out local feature fusion on the input features, and carrying out local feature weight W l Expressed as:
W l ＝sigmoid(F 1 (δ(BN(F 1 (M)))))
further, the output of the local feature fusion branch is expressed as:
Output l ＝I 1 ×W l +I 2 ×(1-W l )
further, the CgT model design includes:
assume that the two-dimensional input feature map isKeys (K), queries (Q) and values (V) are defined respectivelyMeaning k=xw k ，Q＝XW q And v=xw q . Wherein W is q And W is v Is a linear transformation matrix formed by 1 x 1 convolution; w (W) k Is a linear transformation matrix formed by convolution of a group with a convolution kernel size of 3 x 3, and can reflect a static context feature f by context coding an input feature map s At the same time, the auxiliary self-attention learning is used for mining the interaction between the context key features and the query features, and then a dynamic attention weight matrix A is constructed as follows:
A＝g n conv([K,Q])
in the formula g n Conv represents a gated recursive convolution operator.
Then, a dynamic context feature f is calculated d ：
f d ＝A⊙V
By capturing the static context features and the dynamic context features, the following outputs are obtained:
output＝f s +f d 。
another object of the present invention is to provide a photovoltaic hot spot fault detection system applying the photovoltaic hot spot fault detection method, the photovoltaic hot spot fault detection system comprising:
The data acquisition module is used for automatically planning the cruising route of the unmanned aerial vehicle by analyzing the geographic information and the cruising range of the photovoltaic power station, and acquiring a patrol image and a video by utilizing the thermal radiation imaging characteristic of the infrared sensor;
the data transmission module is used for transmitting the inspection data back to the ground control station by virtue of the high-speed low-delay advantage of the 5G wireless network and correspondingly storing the inspection data so as to process the data by a subsequent high-performance computer;
the fault diagnosis module is used for carrying out feature extraction, information aggregation and fault positioning tasks on the photovoltaic data by utilizing a double-branch cooperative training photovoltaic hot spot fault detection algorithm under a designed knowledge distillation mechanism. And finally, combining the image information and GPS (Global Positioning System) positioning data to obtain a hot spot fault diagnosis result.
It is a further object of the present invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the intelligent photovoltaic hot spot fault detection method.
It is a further object of the present invention to provide a computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the intelligent photovoltaic hot spot fault detection method.
The invention further aims at providing an information data processing terminal which is used for realizing the intelligent photovoltaic hot spot fault detection system.
By combining the technical scheme and the technical problems to be solved, the technical scheme to be protected by the invention has the following advantages and positive effects:
first, aiming at the technical problems in the prior art and the difficulty in solving the problems, the technical problems solved by the technical proposal of the invention are analyzed in detail and deeply by tightly combining the technical proposal to be protected, the results and data in the research and development process, and the like, and some technical effects brought after the problems are solved have creative technical effects. The specific description is as follows:
the invention aims to provide a new thought for the photovoltaic hot spot fault detection task, and has a profound effect on accelerating intelligent transformation in the technical field of photovoltaic inspection. Specifically, aiming at the technical problems that the algorithm precision and the model weight are difficult to be compatible in the field of photovoltaic hot spot fault detection at the present stage, the infrared hot spot target characteristics cannot be effectively expressed, the small-scale target detection performance is difficult to ensure, and the network false alarm rate and the false alarm rate are difficult to be reduced by 4, the invention provides a double-branch collaborative training photovoltaic hot spot fault detection system under a knowledge distillation mechanism. Based on knowledge distillation thought, a new mode of teacher and student network collaborative training is designed, the detection accuracy of an algorithm is improved by means of the advantage of a teacher depth detection network, the reasoning efficiency of the algorithm is improved by combining the characteristic of small parameter quantity of the student network, and further the detection accuracy and the requirement of model light weight are considered. The method selects the YOLOX-s algorithm as a student network, greatly reduces the network parameter quantity by constructing an anchor-free frame model, and improves the detection precision of the algorithm and simultaneously accelerates the network convergence speed by combining a decoupling pre-measurement head. A CSPHN module is built in a teacher network to enhance the characteristic expression capability of a backbone network to a hot spot target; designing a BiMAF module to improve the feature aggregation capability of the neck network; and providing a CgT module to further improve the hot spot small-scale target detection performance in various dense scenes, and finally integrating the image information with GPS positioning data to obtain a final hot spot fault diagnosis result. The system combines unmanned aerial vehicle aerial photography and computer vision technology, can autonomously complete large-scale, rapid and fine intelligent hot spot fault detection tasks, and ensures safe and stable operation of a photovoltaic power generation system.
Secondly, the technical scheme is regarded as a whole or from the perspective of products, and the technical scheme to be protected has the following technical effects and advantages:
the invention provides a photovoltaic hot spot fault detection network applicable to complex environments, which carries out cross fusion on an unmanned aerial vehicle inspection technology, a sensor technology and a target detection algorithm and is named as KDBIDet (A bi-branch collaborative training algorithm based on knowledge distillation forphotovoltaic hot spot detection system). The detection network constructs a knowledge distillation module for collaborative training of teachers and students from the viewpoint of model weight reduction so as to improve algorithm reasoning efficiency. In consideration of the positive correlation between target feature extraction and detection performance, a backbone network based on CSPHN is designed, and the feature representation capability of an algorithm is remarkably enhanced under the condition of not consuming excessive network memory by combining the respective advantages of CSPNet and HorNet structures; inspired by the principle of visual nerve layering cognition, a BiMAF model is constructed, and the aggregation capability of a neck network on target features is enhanced by fusing input features from the global and local angles by adopting a parallel method; meanwhile, in order to further enhance the detection performance of the network on the dense small targets, a CgT module is provided before decoupling the prediction head, and an attention matrix is generated by mining static and dynamic context information, so that the detection system can still timely discover hot spot faults existing in the photovoltaic power generation system in a dense small target scene and remove and overhaul the hot spot faults. The network can realize the 'equipment can speak and the power grid can think', and can timely find out and remove and overhaul the hot spot faults existing in the photovoltaic power generation system, so that the safe and stable operation of the photovoltaic power generation system is ensured. With the continuous popularization and application of the product technology, the intelligent transformation of the photovoltaic inspection industry is accelerated, and the transformation and upgrading of the power grid are further promoted.
Thirdly, as inventive supplementary evidence of the claims of the present invention, the following important aspects are also presented:
(1) The expected benefits and commercial values after the technical scheme of the invention is converted are as follows:
according to the invention, a deep learning algorithm is used as a technical support, a cloud-side-end integrated architecture is utilized, and a double-branch collaborative training detection algorithm under a knowledge distillation mechanism is provided for a photovoltaic hot spot fault data set, so that the double-branch collaborative training detection algorithm can be widely applied to various subdivision fields of an intelligent photovoltaic inspection market, and various faults existing in a photovoltaic power generation system are timely discovered and removed and overhauled, so that the safe and stable operation of the photovoltaic power generation system is ensured, and the intelligent photovoltaic power generation system has important research significance and commercial value. In the future, along with the continuous popularization and application of product technology, the intelligent photovoltaic inspection system can realize that equipment can speak and a power grid can think, so that intelligent transformation of the photovoltaic inspection industry is quickened, and power grid transformation and upgrading are further promoted.
(2) The technical scheme of the invention solves the technical problems that people are always desirous of solving but are not successful all the time:
aiming at the technical problems that the algorithm precision and model weight are difficult to be compatible in the field of photovoltaic hot spot fault detection at the present stage, the infrared hot spot target characteristics cannot be effectively expressed, the small-scale target detection performance is difficult to be ensured, the network false alarm rate and the false alarm rate are difficult to be reduced by 4, the invention provides a double-branch cooperative training photovoltaic hot spot fault detection algorithm under a knowledge distillation mechanism, and a knowledge distillation module for cooperative training of teachers and students is constructed from the viewpoint of model weight so as to improve the algorithm reasoning efficiency. From the positive correlation between target feature extraction and detection performance, a backbone network based on CSPHN is designed to enhance the feature expression capability of an algorithm to a target to be detected. Inspired by the principle of visual nerve layering cognition, a BiMAF module is constructed to strengthen the characteristic aggregation capability of a neck network to a multi-scale target. Finally, in order to further restrain adverse effects of target missing false detection on detection precision, a pre-measuring head of a CgT structure is provided, and the model detection precision is improved by executing self-attention learning operation and excavating context information among two-dimensional feature map input keys, so that large-scale, rapid, fine and intelligent operation and maintenance tasks of the photovoltaic power station can be automatically completed. In summary, the present invention can solve the technical problems that need to be solved in the present stage.
(3) The technical scheme of the invention overcomes the technical bias: in the current stage, various intelligent photovoltaic inspection systems generally consider that the algorithm detection precision and the model weight are not compatible, so that in order to ensure the inspection performance, a large-scale detection network with high calculation cost is generally selected to perform fault location and identification on the collected photovoltaic inspection data. Although the accuracy of photovoltaic power station operation maintenance is improved by the mode, the general inspection efficiency is low, and the real-time detection requirement cannot be met. Aiming at the technical prejudice, the invention is inspired by the knowledge distillation idea, and utilizes two modes of local and global distillation to inherit the information in a large teacher network into a compact small student network, so that the detection accuracy of an algorithm is improved by means of the advantage of a teacher depth detection network, the operation speed of the algorithm is accelerated by combining the characteristic of small parameter quantity of the student network, the student network is enabled to obtain strong performance without adding extra cost in the reasoning process, and the large-scale, rapid, fine and intelligent operation and maintenance tasks of a photovoltaic power station are ensured.
Drawings
FIG. 1 is a flow chart of an intelligent photovoltaic hot spot fault detection system provided by an embodiment of the invention;
FIG. 2 is a flowchart of an intelligent photovoltaic hot spot fault detection algorithm provided by an embodiment of the invention;
FIG. 3 is a schematic diagram of an intelligent photovoltaic hot spot fault detection method according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a student detection network provided by an embodiment of the present invention;
FIG. 5 is a diagram of the global distillation principle provided by an embodiment of the present invention;
fig. 6 is a schematic diagram of a CSPHN structure according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of BiMAF structure according to an embodiment of the present invention;
FIG. 8 is a block diagram of an intelligent photovoltaic hot spot fault detection system provided by an embodiment of the present invention;
FIG. 9 is a schematic diagram of three exemplary scenarios provided by embodiments of the present invention; wherein, figure (a) is a small clutter object scene schematic diagram, figure (b) is a noise interference scene schematic diagram, and figure (c) is a motion blur scene schematic diagram;
FIG. 10 is a schematic diagram showing comparison of detection results of different algorithms according to an embodiment of the present invention; wherein, the graph (a) is a schematic diagram of detection results of each algorithm in a small clutter target scene, the graph (b) is a schematic diagram of detection results of each algorithm in a noise interference scene, and the graph (c) is a schematic diagram of detection results of each algorithm in a motion blur scene.
Detailed Description
The present invention will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
In order to fully understand how the invention may be embodied by those skilled in the art, this section is an illustrative embodiment in which the claims are presented for purposes of illustration.
As shown in fig. 1, the intelligent photovoltaic hot spot fault detection system provided by the embodiment of the invention comprises the following steps:
s101, a data acquisition module: the system can automatically plan the cruising route of the unmanned aerial vehicle by analyzing the geographic information and the cruising range of the photovoltaic power station, and collect the inspection image and video by utilizing the thermal radiation imaging characteristic of the infrared sensor;
s102, a data transmission module: by means of the high-speed low-delay advantage of the 5G wireless network, the inspection data are transmitted back to the ground control station and stored correspondingly, so that the subsequent high-performance computer can process the data;
s103, a fault diagnosis module: and carrying out feature extraction, information aggregation and fault positioning tasks on the photovoltaic data by using a designed KDBIDet detection algorithm. Finally, combining the image information and GPS positioning data to obtain a hot spot fault diagnosis result.
As shown in fig. 2, the intelligent photovoltaic hot spot fault detection method provided by the embodiment of the invention comprises the following steps:
s101, constructing a new knowledge distillation collaborative training mode, improving the algorithm detection precision by means of the advantage of a teacher depth network, and improving the algorithm reasoning rate by combining the small parameter quantity characteristics of a student network, so that the detection precision and the model light weight requirements are considered;
S102, selecting a YOLOX detection network for a student network, constructing an anchor-free frame model, thereby greatly reducing the number of network parameters, combining a decoupling prediction head, improving the algorithm detection precision and simultaneously accelerating the network convergence speed, wherein the student network comprises a main network for completing feature extraction, a neck network for performing feature extraction and a decoupling prediction head for executing classification and regression tasks;
s103, constructing a backbone network based on a CSPHB module by a teacher network on the basis of a YOLOX detection network, and obtaining high-order space interaction information similar to a transducer by utilizing a gated convolution and a recursive design on the basis of keeping beneficial induction bias of a convolutional neural network so as to consider the local and global feature expression capability.
S104, a BiMAF module is designed in a teacher network, and multi-level features are aggregated from two angles of global and local in a parallel fusion mode, so that the network can be selectively focused on a region containing target significance information, and further the feature aggregation capability of a neck network is enhanced;
s105, constructing a CgT module before a teacher decouples the prediction head, and improving the expression capability of a transducer architecture by executing self-attention learning operation and excavating context information among input keys of the two-dimensional feature map, so that the hot spot target detection precision under various dense scenes is improved.
As a preferred embodiment, as shown in fig. 3, the method for detecting an intelligent photovoltaic hot spot fault provided by the embodiment of the present invention specifically includes:
1. knowledge distillation cooperative training module
Knowledge distillation aims to inherit information in a large teacher network into a compact student network, so that the information can obtain strong performance without adding extra cost in the reasoning process. However, in the distillation process, the teacher and student networks have imbalance differences between the space and the channel feature patterns, which tend to negatively affect distillation. In response to this problem, the present invention proposes a local-global knowledge distillation structure. It consists of two parts, local and global distillation.
1.1 partial distillation
Aiming at the unbalanced problem of foreground and background characteristics of the teacher and student networks, a local distillation function is designed to separate the foreground and background information of the images and guide the student networks to pay attention to important pixels and channel characteristics.
(1) Construction of binary mask M i,j The image foreground (real frame r) and background are separated as shown in the following formula.
(2) The large-scale targets have larger pixel points and correspondingly larger loss values, so that the distillation is difficult to improve the detection performance of the small-scale targets. Therefore, to balance the loss of the detection network to the targets of different scales and the front background area, a scale mask S is arranged i,j ：
Wherein H is r W and W r The height and width of the target real frame are respectively represented.
(3) In order to make student network concentrate on important space of teacher network and channel information, a space attention mask is constructedChannel attention mask->To improve model distillation performance.
/>
Wherein H, W and C represent feature height, width and channel, respectively; t is a temperature super parameter; f (F) c And F i,j Characteristic information of the c-th channel and characteristic information of size i×j are respectively indicated.
In the training process, a binary mask M is utilized i,j Scale mask S i,j And attention mask A S (F) A is a C (F) The student network is guided to learn the teacher network key space and channel information, so that the following characteristic loss function and attention loss function are constructed:
in the method, in the process of the invention,and->Characteristic diagrams respectively representing teacher and student networks; f (·) is to F S Adjust to F T Reconstruction operators of the same dimension; alpha, beta and gamma are super parameters for balancing various losses; l represents l 1 A norm operator; />And-> The spatial attention mask and the channel attention mask respectively represent the teacher and the student network. The final local distillation function is obtained by calculating the characteristic loss and the attention loss:
L focal ＝L fea +L at (7)
1.2 Global distillation
The local distillation function improves distillation performance by separating the front background information of the image and forcing students to learn the key space and channel characteristics of the teacher network through the network. However, this distillation method cuts off the correlation between the image front and back, resulting in the difficulty of capturing image global information by the detection network, limiting network detection performance. To this end, the present invention proposes a global distillation function to reconstruct the relationship between the different pixels and transfer it from the teacher's network back to the student's network, thereby compensating for the global information lost during the local distillation process, as shown in fig. 4.
The invention captures image global information by utilizing the GcBLock module, thereby setting the following global loss function:
in which W is k ，W v1 And W is v2 Representing a convolution layer; LN representation layer normalization; n (N) p Representing the feature pixel count; lambda represents a balance loss hyper-parameter; r (F) represents feature global information.
In conclusion, the novel knowledge distillation collaborative training mode is constructed, the detection accuracy of the algorithm is improved by means of the advantage of the teacher depth detection network, and the reasoning efficiency of the algorithm is improved by combining the small parameter quantity characteristics of the student network. The Yolox-s is used as a student network, and an improved Yolox-l teacher network is provided so as to meet the requirements of both detection precision and model weight reduction.
2. Student detection network module
The invention selects the YOLOX as a student detection network, greatly reduces the super parameters of manual design by using an anchor-free frame structure, and simultaneously combines a decoupling prediction head to effectively improve the detection precision and accelerate the network convergence speed, as shown in fig. 5. The device consists of a Backbone network (Backbone), a Neck network (Neck) and decoupling pre-measurement heads.
(1) The backbox is mainly composed of three parts, namely Focus, CSPNet and SPP. The Focus slicing module not only enlarges the network receptive field, but also can effectively inhibit the loss of image characteristic information, thereby accelerating the training speed. The CSPNet structure aims to solve the problem that the calculation cost is too high due to repeated gradient information in the network optimization process. The SPP structure expands the backbone network receptive field by using multistage pooling operation, and improves the multi-scale feature fusion capability.
(2) The Neck network builds a CSP2_X structure by referring to the CSPNet to strengthen the network feature fusion capability, and utilizes the FPN+PAN feature pyramid structure to conduct feature aggregation on different scale information. Wherein the FPN conveys strong semantic information top-down and the PAN conveys strong positioning information bottom-up.
(3) And finally, constructing a decoupling pre-measuring head, and decoupling the classifying branch of the focusing texture information and the positioning branch of the focusing edge information, so that the problem of space dislocation is effectively solved, and the convergence speed is improved.
3. Teacher detection network module
3.1 backbone network based on CSPHN Structure
The convolutional neural network has two characteristics of local denaturation, translational denaturation and the like, and the convolutional neural network has higher detection performance even though a small-scale data set is processed by locally focusing on adjacent information in a feature map and adopting the same processing rule for different areas, so that the convolutional neural network has certain limitation on the expression capability of a target global feature in the processes of multiple convolution and pooling operation. Therefore, a transducer model has been developed, which can reflect complex high-order spatial interaction information and long-distance characteristic dependency relationship by constructing a cascade self-attention module, so that the model has excellent global modeling capability, but the performance of the model on a small-scale data set is poor, and secondary calculation complexity exists due to multiple dot product operations.
Therefore, considering the positive correlation between the target feature extraction capability and the detection performance of the backbone network, the invention provides a cross-stage local network based on the HorNet, which obtains high-order space interaction information similar to a Transformer by using gating convolution and recursion design on the basis of keeping the beneficial induction bias of the convolutional neural network, thereby taking the global modeling characteristic into consideration and enhancing the feature expression capability of an algorithm to the target to be detected. The schematic diagram is shown in fig. 6.
As shown in fig. 6, from the view of network structure design, the CSPHN model divides the feature map of the input layer into two parts and combines the two parts by using a cross-stage hierarchical structure, thereby reducing the network memory occupation amount and enhancing the learning ability of the convolutional neural network. Meanwhile, in order to make the backbone network have both local and global feature extraction capability, a residual unit in the original CSPNet model is replaced by a HorNet module. The module inherits the meta-architecture of the transducer model space hybrid layer and feed forward network cascade and utilizes gated recursive convolution (g n Conv) captures the high-order space interaction in the feature map, so that the secondary calculation complexity caused by multiple dot products in the execution process of the multi-head attention mechanism can be avoided, and the extraction capability of the backbone network to the detection target features is improved.
Let g n Conv input features areThen a set of projection features p 0 Is->Can be obtained by the following formula:
wherein phi (·) represents the projection operator, an
Next, a gated recursive convolution is performed by equation (10).
p k+1 ＝DW k (q k )⊙g k (p k )/α,k＝0,1,…,n-1 (10)
Where α is a scaling factor, { DW k The } represents a set of deep convolutions, ++represents a dot-product operation, and { g } is k The dimensions used to match the unused sequences are shown below.
Finally, performing characteristic projection after the top-level recursion operation to obtain g n Output result of Conv. The invention combines the advantages of CSPNet and HorNet to make backbone network have both local and global feature extraction capability. In order to evaluate the superiority of the CSPHN model constructed by the invention, the last layer of CSPNet module of the original Yolox backbone network is replaced by the CSPHN model, and compared with HorNet under the same condition, and the experimental result is shown in Table 1.
Table 1 backbone network comparison results based on HorNet structure and CSPHN structure
As shown in Table 1, the backbone network based on HorNet is subjected to gate-controlled recursive convolution, the number of network parameters and floating point operation times are high, and the CSPHN module constructed by the invention can effectively reduce the network memory occupation amount by constructing a cross-stage hierarchical structure, improve the capturing capability of the backbone network to local and global features, and finally improve the algorithm detection performance.
3.2 neck network based on double-branch multistage characteristic self-adaptive fusion model
The complexity of the background of the captured infrared image is generally high due to the limitation of the remote shooting condition of the unmanned aerial vehicle, and the hot spot fault is mostly a small and medium-sized characteristic distortion target. Thus, to suppress feature differences between multiple scales, the original YOLOX network utilizes a "fpn+pan" structure to pass strong semantic information from top to bottom while passing strong positioning information from top to bottom. However, the structure directly adds and aggregates the feature images with different scales after the feature images with different scales are subjected to size adjustment, so that the trans-scale information of the input end cannot be fully utilized, and the final detection precision is affected.
In order to overcome the defects, the invention designs a dual-branch multistage characteristic self-adaptive fusion module which adopts a parallel method to fuse input characteristics from global and local angles, so that a neck network can be selectively focused on a target area containing significance information, other irrelevant background information is restrained, the detection performance of the model on a small-scale target is improved, and a schematic diagram is shown in figure 7.
(1) Global feature fusion branch
Assume that the multi-stage feature inputs are Input respectively 1 ，The elements and operations are performed as follows.
M＝Input 1 +Input 2 (12)
Global pooling is typically used to globally encode spatial information, but it compresses global spatial information into channel descriptors, making it difficult to preserve feature location information. Thus, to enable the feature fusion module to capture the remote spatial interaction features with precise location information, the global pooling operation is converted into a pair of dimensional feature operators. Specifically, the invention uses two pooling cores with the sizes of (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions respectively, thus obtaining the output of the c-th channel with the height of HAnd the output of the c-th channel with width w +.>
/>
Next, in order to make full use of the feature information obtained after the coordinate information embedding operation, the following concatenation and convolution operations are performed.
f＝δ(F 1 ([z h ,z w ])) (15)
In [ z ] h ,z w ]Representing a join operation along a spatial dimension; delta is the activation function; f is an intermediate feature map obtained by encoding in the horizontal and vertical directions; f represents a 1 x 1 convolution operation.
By decomposing f along the spatial dimension, two independent tensors can be obtained: f (f) h ∈R C/r×H F h ∈R C/r×H Where r is the downsampling ratio. To ensure that the two independent tensors have the same number of channels, the following is performed:
g h ＝F h ([f h ]) (16)
g w ＝F w ([f w ]) (17)
further, global feature weight W g Can be calculated by the following formula.
W g ＝sigmoid(g h )×sigmoid(g w ) (18)
Finally, the global feature fuses the Output of the branch g Can be expressed as:
Output g ＝Input 1 ×W g +Input 2 ×(1-W g ) (19)
(2) Local feature fusion branch
By utilizing the excellent local feature extraction capability of the convolution layer, the invention carries out local feature fusion on the input features and local feature weight W l Can be expressed as:
W l ＝sigmoid(F 1 (δ(BN(F 1 (M))))) (20)
the output of the local feature fusion branch may then be expressed as:
Output l ＝Input 1 ×W l +Input 2 ×(1-W l ) (21)
in summary, the BiMAF model aggregates global and local feature output through parallel modes to obtain a final feature fusion result.
Output＝[Output g ,Output l ] (22)
To verify the superiority of the designed BiMAF model, table 2 gives the comparison results of building only Global feature fusion branches (Global feature fusion, GFF), building only local feature fusion branches (Local feature fusion, LFF), global-local serial feature fusion (Global-local serial feature fusion, GLFF) with the proposed BiMAF model.
TABLE 2 comparison results of multiple feature fusion modules
3.3 prediction head based on CgT model
The original YOLOX detection network adopts an anchor-free framework, so that the number of network parameters can be effectively reduced, and the problem of poor adaptability of a priori anchor frame to a small target is avoided. However, the anchor-free algorithm needs to construct an auxiliary method to acquire a final target detection frame, so that the problems of angle deviation and even target semantic ambiguity are easy to occur, and the detection precision is affected. To overcome the limitations, the invention draws inspiration from the CoT module, builds a g-based model before decoupling the prediction head n The conv CoT model (CgT) improves the expression capability of the transducer architecture by performing self-care learning and mining context information between two-dimensional feature map input keys, thereby improving the hot spot target detection accuracy in various dense scenes.
Specifically, assume that the two-dimensional input feature map isKey (K), query (Q), and value (V) are defined as k=xw, respectively k ，Q＝XW q And v=xw q . Wherein W is q And W is v Is a linear transformation matrix formed by 1 x 1 convolution; w (W) k Is a linear transformation matrix formed by convolution of a group with a convolution kernel size of 3 x 3, and can reflect a static context feature f by context coding an input feature map s At the same time, the auxiliary self-attention learning is used for mining the interaction between the context key features and the query features, and then a dynamic attention weight matrix A is constructed as follows:
A＝g n conv([K,Q]) (23)
in the formula g n Conv represents a gated recursive convolution operator.
In contrast to conventional self-attention weight matrices that rely on isolated key-query pairs, each spatial location of the dynamic attention weight matrix can enhance the feature learning capabilities of the self-attention mechanism by means of static contextual features. The dynamic context feature f can then be calculated d 。
f d ＝A⊙V (24)
By capturing both static and dynamic context features, the following outputs can be obtained:
output＝f s +f d (25)
It should be noted that, although the dot product operation of the CgT module has a problem of secondary complexity, since the input feature map size of the prediction layer is relatively small (20×20,40×40and 80×80), the computational complexity is within a manageable range, and the detection performance is significantly improved.
In summary, in the teacher network, a backbone network based on a CSPHN structure is constructed, a neck network with self-adaptive feature fusion is designed, a prediction head based on a CgT module is provided, and high-performance detection under a high-density small-target scene is realized. Meanwhile, by means of the knowledge distillation idea, important information in a large teacher network is transferred to a compact student network, and finally, a rapid, fine and intelligent photovoltaic hot spot fault detection task is realized.
The intelligent photovoltaic hot spot fault detection system provided by the embodiment of the invention comprises the following components:
under the support of increasingly mature electric power technology, photovoltaic power stations are widely distributed in complex zones with wide regions and sufficient sunlight, such as mountain power stations, water power stations and the like. The coverage area is huge, and the coverage area is generally in a messy and scattered shape due to the limitation of the terrain, so that great challenges are brought to the maintenance work of the power system. In order to solve the problem, the invention provides an intelligent automatic hot spot detection system based on an unmanned plane, and a system frame diagram is shown in fig. 8.
Firstly, the system can automatically plan the cruising route of the unmanned aerial vehicle by analyzing the geographic information and the cruising range of the photovoltaic power station, and collect the inspection image and video by utilizing the thermal radiation imaging characteristic of the infrared sensor. And secondly, by virtue of the advantages of high speed and low time delay of the 5G wireless network, the patrol data are transmitted back to the ground control station and are correspondingly stored, so that the subsequent high-performance computer can process the data. And finally, performing feature extraction, information aggregation and fault positioning tasks on the photovoltaic data by using a designed deep learning algorithm. And finally, combining the image information and the GPS positioning data to obtain a final hot spot fault diagnosis result.
The system aims to perform hot spot fault detection tasks on widely distributed and complex photovoltaic power stations in an autonomous mode. Compared with the traditional manual inspection mode with high operation and maintenance cost, poor working condition and low labor efficiency, the system can complete operation and maintenance work of the photovoltaic module with high precision and high efficiency, timely check out potential safety hazards of the photovoltaic module and remove and overhaul the photovoltaic module, and has important significance for guaranteeing safe and stable operation of the photovoltaic power station.
In order to prove the inventive and technical value of the technical solution of the present invention, this section is an application example on specific products or related technologies of the claim technical solution.
The invention takes a deep learning algorithm as a basis, takes compression, storage, transmission and processing of an image and video as a core, and aims at the complexity and the danger of the geographic position of photovoltaic equipment, the severity of the operation and maintenance problems of a photovoltaic panel and the thermal radiation imaging characteristics of an infrared sensor.
After the system is applied and subjected to the landing, the large-scale, rapid, fine and intelligent operation and maintenance tasks of the photovoltaic power station can be automatically completed, the power generation efficiency of the photovoltaic power station is improved, the operation and maintenance cost of the photovoltaic power station is reduced, the operation and maintenance intellectualization of the photovoltaic power station is realized, meanwhile, the system can be derivative and expanded to be applied to various types of defect detection of the photovoltaic, inspection of substation equipment, multi-target fault detection of a power transmission line and other intelligent construction sites, and the system has important application value.
It should be noted that the embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or special purpose design hardware. Those of ordinary skill in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such as provided on a carrier medium such as a magnetic disk, CD or DVD-ROM, a programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device of the present invention and its modules may be implemented by hardware circuitry, such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., as well as software executed by various types of processors, or by a combination of the above hardware circuitry and software, such as firmware.
The embodiment of the invention has a great advantage in the research and development or use process, and has the following description in combination with data, charts and the like of the test process.
The model training and result analysis provided by the embodiment of the invention are as follows:
in order to evaluate the effectiveness of the proposed algorithm, 500 fault images (640 x 512 in size) are selected from the unmanned aerial vehicle inspection data. Meanwhile, in order to ensure the completeness of a sample, 8 pixels are moved at an angle of 30 degrees in the anticlockwise direction to obtain a motion blurred image, a noise interference image is generated by Gaussian filtering with a variance of 0.2, and 1500 experimental images are finally prepared. According to the intelligent photovoltaic detection system, through carrying out an ablation experiment on each improved module of the algorithm and selecting 7 classical algorithms for qualitative and quantitative comparison analysis, the intelligent photovoltaic detection system is verified to be capable of completing a photovoltaic power station in a large-scale, rapid and fine operation maintenance task under various severe conditions.
1. Ablation experiments
To verify the superiority of each improvement module in the teacher network, ablation experiments were performed on the data set by adding different improvement strategies on the basis of the YOLOX-l detection network, and all experiments used the same data samples and parameter settings, and the comparison results are shown in table 3.
Table 3 comparison results of ablation experiments in teacher network
/>
In Table 3, the CSPHN module is constructed to enhance the backbone network's ability to capture local and global features as compared to the original YOLOX-l detection network. The BiMAF module effectively improves the feature aggregation capability of the neck network by carrying out weight distribution on the multi-scale features. The CgT module combines static and dynamic context information, so that detection performance in dense scenes can be improved. Meanwhile, in order to verify the superiority of cooperative training of different modules, ablation experiments are carried out on a CSPHN+BiMAF network, a CSPHN+ CgT network and a BiMAF+ CgT network, and the results show that compared with an original detection network, the AP index is respectively improved by 0.8%, 1.3% and 1.0%. Finally, by integrating a plurality of improved modules, the detection precision of the research can reach 0.845, the detection performance of the medium-small hot spot targets is obviously improved, and the operation and maintenance tasks of the photovoltaic power station can be completed with high precision even under various severe conditions.
Meanwhile, to evaluate the effectiveness of the knowledge distillation mechanism, the method uses the YOLOX-s algorithm as a student network and uses the improved YOLOX-l algorithm as a teacher network.
Table 4 gives the comparison of ablation experiments before and after introduction of the knowledge distillation mechanism. Experimental results show that the constructed 'teacher+student' collaborative training model can improve the algorithm detection precision while guaranteeing the reasoning efficiency, and further gives consideration to the algorithm detection precision and the model light weight requirement.
Table 4 comparison of ablation experiments before and after introduction of the knowledge distillation mechanism
1. Comparative experiments
In order to objectively evaluate the detection performance of the detection algorithm provided by the invention, 7 detection algorithms of SSD, faster-RCNN, retinanet, FCOS, ATSS, dynamic-RCNN and Yolox are selected for comparison experiments, and the detection results are shown in Table 5.
Table 5 comparison of different detection algorithms
As shown in Table 5, the important information of the teacher network is inherited into the student network by constructing a knowledge distillation model, and the proposed algorithm is applied to AP and AP 50 、AP 75 The index is obviously superior to other 7 comparison algorithms, and the detection performance of the weak hot spot target is obviously improved under the condition of not increasing extra cost. Although compared to the original YOLOX algorithm, the AP M The value is slightly reduced but still near the optimum value.
In order to further prove the superiority of the proposed detection system in various complex environments, the following three typical scenarios are selected for testing and verification: clutter small object scenes, noise interference scenes, and motion blur scenes, as shown in fig. 9. Meanwhile, fig. 10 shows detection results of various algorithms under different scenes. For ease of observation and subsequent analysis, the missed detection and false detection regions of each algorithm have been marked with white solid boxes.
As can be seen from fig. 10 (a), when dealing with a cluttered multi-objective scene, the seven comparison algorithms all have different degrees of omission. In contrast, the KDBIDet system can complete high-precision and high-efficiency photovoltaic module operation and maintenance tasks by designing various improved strategies in a teacher network and transferring important information into a student network by utilizing the knowledge distillation idea.
As can be seen from fig. 10 (b), the SSD, retinanet and YOLOX algorithms have serious missed detection problems due to the interference of complex noise, and the other four comparison detection networks have insufficient characteristic information aggregation capability on different scales, so that the multi-scale hot spot target accurate detection task is difficult to realize. The constructed KDBIDet network aggregates multi-scale features from two angles of global and local in a parallel fusion mode, so that the network can be selectively focused on a region containing target significance information, and further the trans-scale feature aggregation capability of an algorithm is enhanced.
As can be seen from fig. 10 (c), in the motion blur scene, the infrared image cannot accurately express the outline feature of the hot spot fault, so that the detection performance of the algorithm under the condition of dense multiple targets is limited. The designed KDBIDet improves the expression capability of a transducer architecture by executing self-attention learning operation and mining context information among input keys of the two-dimensional feature map, thereby improving the detection precision of hot spot small-scale targets in various dense scenes.
The foregoing is merely illustrative of specific embodiments of the present invention, and the scope of the invention is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present invention will be apparent to those skilled in the art within the scope of the present invention.

Claims (10)

1. The intelligent photovoltaic hot spot fault detection method is characterized by comprising the following steps of:
constructing a knowledge distillation module for cooperative training of a teacher and a student; designing a backbone network enhancement algorithm based on CSPHN, and expressing the characteristic of the target to be detected; constructing a BiMAF model, and fusing input features from the global and local angles by a parallel method to strengthen the aggregation capability of a neck network on target features; meanwhile, a CgT module is provided for finding out and removing and overhauling the hot spot faults existing in the photovoltaic power generation system under various severe environments.
2. The intelligent photovoltaic hot spot fault detection method according to claim 1, wherein the intelligent photovoltaic hot spot fault detection method comprises the steps of:
firstly, constructing a new mode of cooperative training of a teacher and a student network based on a knowledge distillation idea, improving the detection accuracy of an algorithm by means of the advantage of a teacher depth detection network, and improving the reasoning efficiency of the algorithm by combining the small parameter quantity characteristics of the student network, thereby taking into account the detection accuracy and the light weight requirement of a model;
Secondly, building a student network based on a YOLOX algorithm based on an unmanned aerial vehicle intelligent inspection data set, introducing an anchor-free frame structure, directly predicting the position coordinates of a target through key points without setting a specific anchor frame, and simultaneously combining a decoupling prediction head so as to improve the detection precision of the algorithm and accelerate the convergence speed of the network;
thirdly, constructing a backbone network based on a CSPHN model in a teacher network, and obtaining high-order space interaction information similar to a Transformer by utilizing gated convolution and recursive design on the basis of keeping beneficial induction bias of a convolutional neural network so as to enhance the characteristic expression capability of an algorithm on a hot spot target;
step four, a BiMAF (Bi-branch multi-level feature adaptive fusion) module is applied to the neck network, and multi-level features are aggregated from two angles of global and local in a parallel fusion mode, so that the network can be selectively focused on a region containing target significance information, and the feature aggregation capability of the neck network is further enhanced;
and fifthly, constructing a decoupling pre-measurement head based on a CgT module in a teacher network, generating an attention matrix by mining static and dynamic context information, so that a detection system can still timely find out hot spot faults existing in a photovoltaic power generation system in a dense small target scene, and removing and overhauling the hot spot faults.
3. The intelligent photovoltaic hot spot fault detection method according to claim 2, wherein the constructing of the co-training new pattern comprises:
(1.1) designing a local distillation function to separate image foreground and background information, and guiding a student network to pay attention to important pixels and channel characteristics;
(1.2) providing a global distillation function to reconstruct the relationship between different pixels and transmitting the relationship between different pixels from the teacher network back to the student network for compensating the global information lost in the local distillation process;
(1.3) inheriting information in the teacher network into the student network by fusing local distillation and global distillation.
4. The intelligent photovoltaic hot spot fault detection method according to claim 3, wherein the specific process of step (1.1) comprises:
(1.1.1) constructing a binary mask M for a feature map F having an abscissa of i and an ordinate of j, respectively i,j Separating the foreground and background of the image, as shown in the following formula:
wherein r is a target real frame.
(1.1.2) setting a scale mask S for balancing the loss of the detection network to different scale targets and foreground areas i,j ：
Wherein H is r W and W r Respectively representing the height and the width of a target real frame;
(1.1.3) construction of a spatial attention mask Channel attention mask->To improve model distillation performance:
wherein H, W and C represent feature height, width and channel, respectively; t is a temperature super parameter; f (F) c And F i,j Characteristic information of a c-th channel and characteristic information of a size of i×j are respectively represented;
(1.1.4) during training, binary mask M is used i,j Scale mask S i,j And attention maskIs->Guiding students to learn teacher network key space and channel information through network, thereby constructing the following characteristic loss function L fea Attention loss function L at ：
In the method, in the process of the invention,and->Characteristic diagrams respectively representing teacher and student networks; f (·) is to F S Adjust to F T Reconstruction operators of the same dimension; alpha, beta and gamma are super parameters for balancing various losses; l represents l 1 A norm operator; />And->Respectively representing a spatial attention mask and a channel attention mask corresponding to the teacher and the student network;
(1.1.5) obtaining a final local distillation function by calculating the characteristic loss and the attention loss:
L focal ＝L fea +L at
5. the intelligent photovoltaic hot spot fault detection method according to claim 3, wherein the specific process of step (1.2) comprises:
capturing global information of an image by using a GcBLock module so as to obtain global loss L global ：
L global ＝λ·∑(R(F T )-R(F S )) 2
Wherein λ represents a balance loss hyper-parameter; r (F) represents feature global information, which can be expressed as:
In which W is k ，W v1 And W is v2 Representing a convolution layer; LN representation layer normalization; n (N) p Representing the feature pixel count; reLU represents a linear activation function.
6. The intelligent photovoltaic hot spot fault detection method according to claim 2, wherein the designing of the student network comprises:
the YOLOX network includes a Backbone network (Backbone), a Neck network (neg) and decoupling pre-measurement heads;
the backbox includes Focus slice modules, CSPNet (Cross Stage Partial Network) and SPP (Spatial Pyramid Pooling); the Focus slicing module is used for expanding a network receptive field and inhibiting image characteristic information loss; the CSPNet structure is used for solving the problem that the calculation cost is too high due to repeated gradient information in the network optimization process; the SPP structure expands a backbone network receptive field by utilizing multi-stage pooling operation;
the Neck is composed of a CSP2_X structure, and utilizes a FPN+PAN feature pyramid structure to perform feature aggregation on different scale information, wherein the FPN transmits strong semantic information from top to bottom, and the PAN transmits strong positioning information from bottom to top;
the decoupling pre-measurement head is used for decoupling the classification branch of the focusing texture information and the positioning branch of the focusing edge information.
7. The intelligent photovoltaic hot spot fault detection method according to claim 2, wherein the constructing of the CSPHN module includes:
dividing the feature map of the input layer into two parts, and combining the two parts by utilizing a cross-stage hierarchical structure; replacing residual units in the original CSPNet model with HorNet modules inheriting the meta-architecture of the Transformer model space hybrid layer and feed-forward network cascade and utilizing g n Conv captures the higher-order spatial interactions in the feature map;
the specific process comprises the following steps:
first, let g n Conv input features areThen a set of projection features p 0 Is->Can be obtained by the following formula:
wherein phi (·) represents the projection operator, an
Then, a gated recursive convolution is performed, the formula:
p k+1 ＝DW k (q k )⊙g k (p k )/α,k＝0,1,...,n-1
where α is a scaling factor, { DW k The } represents a set of deep convolutions, { g, represents a dot product operation k The dimensions used to match the unused sequences are shown as follows:
finally, performing feature projection after the top-level recursion operation to obtain g n Output result of Conv.
8. The intelligent photovoltaic hot spot fault detection method according to claim 2, wherein the specific process of the BiMAF module comprises:
(2.1) global feature fusion tributaries: converting global pooling operation into one-to-one dimensional feature operators, outputting global feature weights through cascading and convolution operation and utilizing sigmoid normalization function, and obtaining global feature fusion result Output through weight distribution operation g ；
(2.2) constructing local feature weights by using a 1X 1 convolution and sigmoid normalization function, and further obtaining a local feature fusion result Output by using weight distribution operation l ；
(2.3) aggregating the global feature fusion result and the local feature fusion result by the BiMAF model in a parallel fusion mode to obtain a final feature fusion result:
Output＝[Output g ,Output l ]。
9. the intelligent photovoltaic hot spot fault detection method according to claim 2, wherein the calculation process of the global feature fusion result in step (2.1) includes:
(2.1.1) assume that the multi-level feature inputs are Input, respectively 1 ，The elements and operations are performed as follows:
M＝Input 1 +Input 2
encoding each channel in the horizontal and vertical directions by using two pooling cores with the sizes of (H, 1) and (1, W) to obtain the output of the c-th channel with the height of HAnd the output of the c-th channel with width w +.>
(2.1.2) performing the following concatenation and convolution operations:
f＝δ(F 1 ([z h ,z w ]))
in [ z ] h ,z w ]Representing a join operation along a spatial dimension; delta is the activation function; f is an intermediate feature map obtained by encoding in the horizontal and vertical directions; f represents a 1×1 convolution operation;
by decomposing f along the spatial dimension, we getTwo independent tensors: f (f) h ∈R C/r×H F h ∈R C/r×H Wherein r is the downsampling ratio; at the same time, two 1X 1 convolutions F are performed h And F w Operate to ensure that two independent tensors have the same number of channels, thereby obtaining global feature weights W g ：
W g ＝sigmoid(F h ([f h ]))×sigmoid([f w ])
(2.1.3) Output of Global feature fusion branch g The method comprises the following steps:
Output g ＝I 1 ×W g +I 2 ×(1-W g )
the calculation process of the local feature fusion result in the step (2.2) comprises the following steps:
carrying out local feature fusion on the input features, and carrying out local feature weight W l Expressed as:
W l ＝sigmoid(F 1 (δ(BN(F 1 (M)))))
further, the output of the local feature fusion branch is expressed as:
Output l ＝Input 1 ×W l +Input 2 ×(1-W l )；
the specific process of the CgT model comprises the following steps:
assume that the two-dimensional input feature map isKey (K), query (Q), and value (V) are defined as k=xw, respectively k ，Q＝XW q And v=xw q . Wherein W is q And W is v Is a linear transformation matrix formed by 1 x 1 convolution; w (W) k Is a linear transformation matrix formed by convolution of a group with a convolution kernel size of 3 x 3, and can reflect a static context feature f by context coding an input feature map s At the same time, the auxiliary self-attention learning is used for mining the interaction between the context key features and the query features, and then a dynamic attention weight matrix A is constructed as follows:
A＝g n conv([K,Q])
in the formula g n Conv represents a gated recursive convolution operator.
Then, a dynamic context feature f is calculated d ：
f d ＝A⊙V
By capturing the static context features and the dynamic context features, the following outputs are obtained:
output＝f s +f d 。
10. Photovoltaic hot spot fault detection implementing the photovoltaic hot spot fault detection method according to any one of claims 1-10, characterized in that the photovoltaic hot spot fault detection system comprises:
the data acquisition module is used for automatically planning the cruising route of the unmanned aerial vehicle by analyzing the geographic information and the cruising range of the photovoltaic power station, and acquiring a patrol image and a video by utilizing the thermal radiation imaging characteristic of the infrared sensor;
the data transmission module is used for transmitting the inspection data back to the ground control station by virtue of the high-speed low-delay advantage of the 5G wireless network and correspondingly storing the inspection data so as to process the data by a subsequent high-performance computer;
and the fault diagnosis module is used for carrying out feature extraction, information aggregation and fault positioning tasks on the photovoltaic data by utilizing a double-branch cooperative training photovoltaic hot spot fault detection algorithm under a knowledge distillation mechanism, and finally combining the image information and GPS positioning data to obtain a hot spot fault diagnosis result.
CN202310028596.5A 2023-01-09 2023-01-09 Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal Pending CN117095311A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310028596.5A CN117095311A (en) 2023-01-09 2023-01-09 Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310028596.5A CN117095311A (en) 2023-01-09 2023-01-09 Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal

Publications (1)

Publication Number Publication Date
CN117095311A true CN117095311A (en) 2023-11-21

Family

ID=88780619

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310028596.5A Pending CN117095311A (en) 2023-01-09 2023-01-09 Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal

Country Status (1)

Country Link
CN (1) CN117095311A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274723A (en) * 2023-11-22 2023-12-22 国网智能科技股份有限公司 Target identification method, system, medium and equipment for power transmission inspection

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274723A (en) * 2023-11-22 2023-12-22 国网智能科技股份有限公司 Target identification method, system, medium and equipment for power transmission inspection
CN117274723B (en) * 2023-11-22 2024-03-26 国网智能科技股份有限公司 Target identification method, system, medium and equipment for power transmission inspection

Similar Documents

Publication Publication Date Title
Jenssen et al. Intelligent monitoring and inspection of power line components powered by UAVs and deep learning
Jiao et al. Msmdfusion: Fusing lidar and camera at multiple scales with multi-depth seeds for 3d object detection
Wang et al. Research on image recognition of insulators based on YOLO algorithm
CN117095311A (en) Intelligent photovoltaic hot spot fault detection method, system, medium, equipment and terminal
CN114021741A (en) Photovoltaic cell panel inspection method based on deep learning
CN114596278A (en) Method and device for detecting hot spot defects of photovoltaic panel of photovoltaic power station
CN115131747A (en) Knowledge distillation-based power transmission channel engineering vehicle target detection method and system
CN115223009A (en) Small target detection method and device based on improved YOLOv5
Yin et al. PV-YOLO: lightweight YOLO for photovoltaic panel fault detection
Manninen et al. Multi-stage deep learning networks for automated assessment of electricity transmission infrastructure using fly-by images
Cao et al. Improved YOLOv8-GD deep learning model for defect detection in electroluminescence images of solar photovoltaic modules
CN117237696A (en) Substation weed detection method and system based on YOLOv5 and attention mechanism
Wang et al. YOLO-ERF: lightweight object detector for UAV aerial images
CN113496210B (en) Photovoltaic string tracking and fault tracking method based on attention mechanism
Liquan et al. Fast detection of defective insulator based on improved YOLOv5s
Ruan et al. Deep Learning-based Method for PV Panels Segmentation and Defects Detection with Infrared Images
CN114821224A (en) Method and system for amplifying railway image style conversion data
Sato et al. Semantic Segmentation of Outcrop Images using Deep Learning Networks Toward Realization of Carbon Capture and Storage
Xia et al. Application of yolov7 and transformer structures to small object (license plate) detection in complex traffic scenes
Prabhakaran et al. Defect analysis of faulty regions in photovoltaic panels using deep learning method
Hu et al. Two-stage insulator self-explosion defect detection method based on Mask R-CNN
Zhang et al. Research on steel surface defect detection based on YOLOv5
Mu et al. UAV image defect detection method for steel structure of high-speed railway bridge girder
Wang et al. Photovoltaic Panel Intelligent Detection Method Based on Improved Faster-RCNN
Pengcheng et al. A RealTime Image Recognition Method of Power AI Based on Quadtree Algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination