CN114973032A - Photovoltaic panel hot spot detection method and device based on deep convolutional neural network - Google Patents
Photovoltaic panel hot spot detection method and device based on deep convolutional neural network Download PDFInfo
- Publication number
- CN114973032A CN114973032A CN202210587259.5A CN202210587259A CN114973032A CN 114973032 A CN114973032 A CN 114973032A CN 202210587259 A CN202210587259 A CN 202210587259A CN 114973032 A CN114973032 A CN 114973032A
- Authority
- CN
- China
- Prior art keywords
- photovoltaic panel
- hot spot
- network
- infrared image
- model
- 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 59
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 20
- 230000011218 segmentation Effects 0.000 claims abstract description 62
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 53
- 238000000605 extraction Methods 0.000 claims abstract description 46
- 238000000034 method Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims description 44
- 230000006870 function Effects 0.000 claims description 34
- 238000012360 testing method Methods 0.000 claims description 21
- 230000008014 freezing Effects 0.000 claims description 12
- 238000007710 freezing Methods 0.000 claims description 12
- 238000010257 thawing Methods 0.000 claims description 10
- 230000006872 improvement Effects 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 8
- 238000013434 data augmentation Methods 0.000 claims description 6
- 238000005516 engineering process Methods 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000006835 compression Effects 0.000 claims description 3
- 238000007906 compression Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 description 12
- 238000005070 sampling Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 5
- 238000012986 modification Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 4
- 238000011897 real-time detection Methods 0.000 description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 3
- 230000003321 amplification Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 229910000906 Bronze Inorganic materials 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000003042 antagnostic effect Effects 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000010974 bronze Substances 0.000 description 1
- KUNSUQLRTQLHQQ-UHFFFAOYSA-N copper tin Chemical compound [Cu].[Sn] KUNSUQLRTQLHQQ-UHFFFAOYSA-N 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 238000004146 energy storage Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 210000003608 fece Anatomy 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 235000020068 maotai Nutrition 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000002715 modification method Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000001931 thermography Methods 0.000 description 1
Images
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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- 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
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/50—Photovoltaic [PV] energy
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biodiversity & Conservation Biology (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a photovoltaic panel hot spot detection method and device based on a deep convolutional neural network, belonging to the technical field of photovoltaic panel hot spot detection and comprising the following steps: acquiring an infrared image of the photovoltaic panel; constructing a photovoltaic panel identification model by using an improved target detection algorithm Yolov4, and identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model; and constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting the hot spots on the intercepted photovoltaic panel through the hot spot segmentation model. According to the method, the Yolov4 feature extraction network is replaced, the photovoltaic panel of the aerial infrared image is rapidly identified, the problem of influence of the ground background of the infrared image is solved, the MobileNet V2 network is introduced into the Deeplab V3+ model to rapidly segment the hot spot of the identification result, and the hot spot can be accurately detected.
Description
Technical Field
The invention belongs to the technical field of hot spot detection of photovoltaic panels, and particularly relates to a hot spot detection method and device of a photovoltaic panel based on a deep convolutional neural network.
Background
With the continuous promotion of new energy, photovoltaic power stations are connected to rural power grids on a large scale. In recent years, the cost of the photovoltaic industry is continuously reduced, and the productivity is rapidly increased. The photovoltaic power station is also an important measure for increasing the income of the poverty-stricken users and benefiting poverty-stricken areas. However, in view of the rural ecological environment, the photovoltaic panel of the rural distributed power station is very easy to be shielded by dust, bird droppings and fallen leaves, and if the photovoltaic panel cannot be cleaned in time, the shielded cells become loads to consume energy, and hot spot faults are generated. The hot spot fault greatly reduces the generating efficiency of the photovoltaic panel, and even damages the whole photovoltaic panel and even has fire hazard. Therefore, the hot spot detection of the photovoltaic panel has important value for promoting the photovoltaic energy storage work.
At present, the mode of detecting the defects of the photovoltaic panel mainly comprises image processing, thermal imaging, neural network and the like. Tsanakas et al use a conventional image processing algorithm to detect hot spot failures, but need to manually segment the background of the image; the Wangbezhen and the like propose a method for detecting hot spot faults by analyzing infrared images aiming at the characteristic that the solar cells in different working states have different working temperatures, but the detection result is greatly influenced by the environment; the Sunhua and the like are used for identifying and positioning the photovoltaic hot spots of the small samples by using a deep convolution self-coding network model, and a deep migration learning model is constructed on the basis of an inclusion-v 3 model, but the two methods are only suitable for detecting the hot spots of the small samples; wangchun proposed generating an antagonistic neural network for hot spot identification, but not applicable to situations where an incomplete photovoltaic panel assembly is present; the method is characterized in that the photovoltaic array is segmented by local gray scale features of a photovoltaic array region for the old civilization, and then hot spot detection is carried out by adopting an SVM (support vector machine), but the model is longer in training time and is not suitable for real-time detection of an unmanned aerial vehicle.
Disclosure of Invention
Aiming at overcoming the defects of the prior art, the invention provides a photovoltaic panel hot spot detection method and device based on a deep convolutional neural network by combining an unmanned aerial vehicle inspection technology, aiming at the problem that hot spot faults of photovoltaic panels of rural distributed photovoltaic power stations in China are difficult to detect. In view of the problems and the actual requirement of real-time detection of the unmanned aerial vehicle, the invention replaces the Yolov4 feature extraction network, realizes the quick identification of the aerial infrared image photovoltaic panel, and solves the problem of the influence of the ground background of the infrared image. And then, the identified infrared image is subjected to hot spot detection by using a Deeplabv3+ semantic segmentation algorithm for replacing the feature extraction network, so that the hot spot can be accurately detected.
In order to achieve the above purpose, the invention provides the following technical scheme:
a photovoltaic panel hot spot detection method based on a deep convolutional neural network comprises the following steps:
acquiring an infrared image of the photovoltaic panel;
constructing a photovoltaic panel identification model by using an improved target detection algorithm Yolov4, and identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model; the photovoltaic panel can be rapidly identified from the infrared image;
constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting hot spots on the intercepted photovoltaic panel through the hot spot segmentation model; the fast segmentation of the hot spots of the identification result is realized, and the target loss caused by down sampling is improved;
the improvement of the target detection algorithm Yolov4 comprises replacing the main feature extraction network CSPDarknet53 of the Yolov4 algorithm with a lightweight network MobileNet V2, and replacing the standard 3 × 3 convolution in the enhanced feature extraction network PANet with a deep separable convolution; the optimization of the semantic segmentation algorithm DeepLabV3+ includes replacing the backbone feature extraction network Xconcept with a lightweight network MobileNetV 2.
Preferably, the photovoltaic panel infrared image is acquired by unmanned aerial vehicle aerial photography.
Preferably, identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model comprises the following steps:
performing primary feature extraction on the input photovoltaic panel infrared image through a main feature extraction network MobileNet V2 of the improved target detection algorithm Yolov4 to obtain a primary feature layer;
pooling processing of different scales is carried out on the characteristic layer through a spatial pyramid pooling network SPP of the improved target detection algorithm Yolov4, and after cascading, 3x3 convolution is carried out;
inputting the convolved features into an enhanced feature extraction network (PAne) for feature fusion to obtain four coordinates top, left, bottom and right of a prediction frame;
and identifying the photovoltaic panel identification model by using a matrix form through the four coordinates top, left, bottom and right of the obtained prediction frame to obtain a photovoltaic panel identification result picture and intercepting the photovoltaic panel identification result picture.
Preferably, the hot spot detection and segmentation of the intercepted hot spot on the photovoltaic panel is performed through the hot spot segmentation model, and the method comprises the following steps:
in an encoder of the optimized semantic segmentation algorithm DeepLabV3+, extracting hot spot features on a photovoltaic panel by using a cavity convolution of a MobileNet V2 network, merging the hot spot features, performing 1 × 1 convolution compression, and outputting advanced features;
in a decoder of the optimized semantic segmentation algorithm DeepLabV3+, the low-level features in the encoder are subjected to 1 × 1 convolution dimensionality reduction, then are subjected to feature fusion with the high-level features, the target boundary is restored, and the hot spots on the final photovoltaic panel are decoded.
Preferably, the loss function adopted by the optimized semantic segmentation algorithm deep labv3+ is a loss function obtained by combining a Dice loss function and a cross entropy loss function, and the expression is as follows:
wherein H is the height of the image, W is the width of the image, p i，j Is the predicted probability, g, of having a foreground at position (i, j) i，j The label at position (i, j) is 0 or 1.
Preferably, the photovoltaic panel infrared image is preprocessed through a data augmentation technology before the photovoltaic panel identification is performed through the photovoltaic panel identification model, so that data expansion is realized.
Preferably, the photovoltaic panel infrared image data is preprocessed by using a data augmentation technology, and the method comprises the following steps:
randomly rotating the photovoltaic panel infrared image by 0-120 degrees to form an expanded photovoltaic panel infrared image;
or translating the photovoltaic panel infrared image to the right and rotating the photovoltaic panel infrared image by 180 degrees to form an expanded photovoltaic panel infrared image so that the detection target is positioned at different positions;
or the infrared image of the photovoltaic panel is darkened to form an expanded infrared image of the photovoltaic panel, so that the problem occurring in actual shooting is simulated;
and marking the expanded photovoltaic panel infrared image for network training.
Preferably, the method further comprises the step of performing parameter training on the photovoltaic panel identification model and the hot spot segmentation model by using the augmented photovoltaic panel infrared image, and the parameter training comprises the following steps:
the parameter training of the photovoltaic panel recognition model comprises:
and (3) the expanded photovoltaic panel infrared image is processed according to the proportion of a training set to a verification set of 8: 2, dividing the training process into a freezing stage and a thawing stage;
in the freezing stage, the main network is frozen, and the feature extraction network is not changed; in the unfreezing stage, the trunk is not frozen, and the feature extraction network is changed;
using the trained pre-training weight, training 50 times with the initial learning rate of 1e-3 in the freezing stage, and training 50 times with the initial learning rate of 1e-4 in the thawing stage, wherein the total number of the training times is 100; after the trunk network is replaced, each model is subjected to repeated tests for 3 times, and the average value is taken as a test result;
the parameter training of the hot spot segmentation model comprises:
screening out a plurality of pictures with hot spots in the pictures of the identification results of the photovoltaic panel according to the proportion of a training set to a verification set 9: 1, dividing;
using the trained pre-training weight, wherein the initial learning rate is 1e-3 in the freezing stage and is trained for 35 times, and the initial learning rate is 1e-4 in the thawing stage and is trained for 35 times, and the training times are 70 times; the results were also obtained by averaging the respective indices in 3 trials.
Another object of the present invention is to provide a photovoltaic panel hot spot detection apparatus based on a deep convolutional neural network, including:
the image acquisition unit is used for acquiring an infrared image of the photovoltaic panel;
the identification unit is used for constructing a photovoltaic panel identification model by utilizing an improved Yolov4 algorithm, identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model;
the segmentation unit is used for constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting hot spots on the intercepted photovoltaic panel through the hot spot segmentation model;
the improvement of the target detection algorithm Yolov4 comprises replacing the main feature extraction network CSPDarknet53 of the Yolov4 algorithm with a lightweight network MobileNet V2, and replacing the standard 3 × 3 convolution in the enhanced feature extraction network PANet with a deep separable convolution; the optimization of the semantic segmentation algorithm DeepLabV3+ includes replacing the backbone feature extraction network Xconcept with a lightweight network MobileNetV 2.
The photovoltaic panel hot spot detection method and device based on the deep convolutional neural network have the following beneficial effects:
according to the method, a Yolov4 trunk feature extraction network is replaced by a lightweight network MobileNet V2, a standard 3 multiplied by 3 convolution in a PANet network is replaced by a depth separable convolution, and the Yolov4 feature extraction network is replaced to realize quick identification of the aerial infrared image photovoltaic panel, realize quick identification of the photovoltaic panel from the infrared image and solve the problem of influence of the ground background of the infrared image; the rapid segmentation of the hot spots of the recognition result is realized by introducing a MobileNetV2 network into a Deeplab V3+ model, so that the hot spots can be accurately detected.
Drawings
In order to more clearly illustrate the embodiments of the present invention and the design thereof, the drawings required for the embodiments will be briefly described below. The drawings in the following description are only some embodiments of the invention and it will be clear to a person skilled in the art that other drawings can be derived from them without inventive effort.
Fig. 1 is a flowchart of a photovoltaic panel hot spot detection method based on a deep convolutional neural network according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a depth separable convolution decomposition;
FIG. 3 is a diagram of a reverse residual structure;
FIG. 4 is a network structure diagram of the mobilenetv2-yolov4-lite algorithm;
FIG. 5 is a diagram of a Deeplab V3+ _ MobileNetV2 network architecture;
FIG. 6 is a raw image acquired for a test;
FIG. 7 is a diagram showing the result of the infrared image amplification;
FIG. 8 is a graph of the loss function under the MobileNet V2-Yolov4-lite model;
FIG. 9 is a photovoltaic panel recognition result diagram under a MobileNet V2-Yolov4-lite model;
FIG. 10 is a loss function curve under a Deeplab V3+ _ MobileNet V2 modified downsampling and loss function model;
FIG. 11 is a graph of the hot spot segmentation results under the modified downsampling and loss function model of Deeplab V3+ _ MobileNet V2.
Detailed Description
In order that those skilled in the art will better understand the technical solutions of the present invention and can practice the same, the present invention will be described in detail with reference to the accompanying drawings and specific examples. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example 1
The invention provides a photovoltaic panel hot spot detection method based on a deep convolutional neural network, which comprises the following steps of:
Step 1.1, acquiring infrared images of photovoltaic panels
As shown in fig. 6, image data is collected from photovoltaic power stations in the new autumn town after the dawn county, maotai, new fun, and the XT2 tripod head camera is carried by a longitude and latitude M300 drone in dayang to collect infrared images of a photovoltaic panel with a height of 30 meters, 2188 infrared images are collected in total, and 1557 infrared images containing the photovoltaic panel are screened out for network training and testing.
Step 1.2, preprocessing the infrared image of the photovoltaic panel
In order to solve the problems of insufficient network training samples and insufficient diversity, a data augmentation technology is required to reduce the risk of network overfitting. The method is characterized in that an original image is randomly rotated (0-120 degrees) to form an expanded photovoltaic panel infrared image, the photographed photovoltaic panel images are consistent due to the design problem of a photographing route of the unmanned aerial vehicle, and the problem of sample diversity can be solved through random angle rotation. The second method is to translate and rotate the image by 180 ° to the right, so that the detection target is located at a different position. A third method is to darken the image to simulate some problems that may occur when actually shooting.
After data expansion, 6228 photovoltaic panel infrared images are obtained, marked and used for network training, and the data amplification effect is shown in fig. 7.
Step 2, constructing a photovoltaic panel identification model by using an improved Yolov4 algorithm, identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model;
yolov4 is based on a great deal of previous research, and by combining and innovating algorithms, the detection speed is improved and the accuracy is ensured. The Yolov4 network structure mainly comprises a main feature extraction network CSPDarknet53, a spatial pyramid pooling network SPP, a reinforced feature extraction network PANet and a prediction network YOLO-Head. The CSPDarknet53 is activated by a Mish activation function, and the feature extraction capability and the network stability of the backbone network are improved. The representation of the function Mish is (1) - (4):
Mish＝x×tanh[ln(1+e x )] (1)
the loss function used in Yolov4 is CIoU-loss, which is calculated as:
v is the similarity of aspect ratio; c is the lower graph diagonal distance; alpha is a weight parameter.
The MobileNet network requires less computing resources, is suitable for mobile equipment, and uses a deep separable convolution formed by deep convolution and point-by-point convolution instead of standard convolution. Wherein, the depth convolution filter can filter the input channels on the basis of not increasing the number of channels, and the point-by-point convolution filter can increase the number of channels. This results in a significant reduction in the number of computations and can be arranged at the mobile end with a depth separable convolution decomposition as shown in figure 2.
MobileNetV2 is an upgrade based on MobileNetV1, the use of an Inverted residual block (Inverted residual block) is an important improvement, and the entire MobileNetV2 is composed of an Inverted residual block, which can be divided into two parts, as shown in fig. 3: on the left is the stem part, compared to MobileNetV1, which is first convolved by 1x1 for upscaling to obtain more features, then convolved by 3x3 depth separable convolution, and finally convolved by 1x1 for downscaling. The right side is the residual side part, and the input and the output are directly connected.
In the embodiment, the photovoltaic panel identification model is obtained by improving the target detection algorithm Yolov4, and the improvement on the Yolov4 algorithm mainly has the following two aspects.
Firstly, a Yolov4 trunk feature extraction network CSPDarknet53 is replaced by using a MobileNet V1 network, a MobileNet V2 network and a MobileNet V3 network respectively, the problems of large parameter quantity and large calculated quantity in CSPDarknet53 network feature extraction are solved while the precision is ensured, and the MobileNet V2 network which shows the best performance in the data set of the invention is preferably selected as the feature extraction network of the model by comparing the speed, the detection precision, the model size and the like. Secondly, the standard 3 × 3 convolution in the enhanced feature extraction network PANet is replaced by a deep separable convolution with a smaller calculation parameter amount, so that the calculation parameter amount is further reduced, and the improved algorithm network structure is shown in fig. 4.
The function of the Backbone feature extraction network Backbone is to perform preliminary feature extraction on an input 416 × 416 × 3 picture, so as to obtain three preliminary effective feature layers with shape of 13 × 13, 26 × 26, and 52 × 52, which contain semantic information with different dimensions. Therefore, the three effective feature layers can be replaced before the reinforced feature extraction network is input, so that the purpose of replacing the network is achieved.
The enhanced feature extraction network corresponds to the Spatial Pyramid Pooling (SPP) and aggregation network (PANet) shown in fig. 4, and firstly, the SPP performs pooling processing of different scales on input 13 × 13 feature layers, performs 3 × 3 convolution after cascading, and finally enters the PANet to perform feature fusion with 26 × 26 and 52 × 52 feature layers, so as to extract better features for output. To further reduce the number of parameters, we can use a 3 × 3 deep separable convolution plus a 1 × 1 standard convolution instead of the standard 3 × 3 convolution used by PANet in Yolov4, since its parameters are mainly focused on the standard 3 × 3 convolution.
And finally, intercepting the recognition result of the photovoltaic panel by acquiring four coordinates top, left, bottom and right of the prediction frame in a matrix form and placing the result under a black screen.
And 3, constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting the hot spots on the intercepted photovoltaic panel through the hot spot segmentation model.
The DeepLabV3+ segmentation model is based on the DeepLabV3 and adds a decoder module to obtain clearer segmentation results. An encoder part (encoder) of the DeepLabV3+ extracts image characteristic information through depth separable convolution layers of different channels in a backbone network Xconcentration model, and performs characteristic extraction by utilizing Spatial Pyramid Pooling (SPP). Low-level features are fused with high-level features in a decoder (decoder). The decoder and the SPP both use the depth separable convolution instead of the hole convolution, and reduce the calculation parameters, thereby realizing a faster and stronger codec network.
Cross entropy is often used in Deeplabv3+ as a loss function, calculated as (5):
where H is the height of the image, W is the width of the image, p ,i,j Is the predicted probability at position (i, j) corresponding to the label at that position. p is a radical of ,i,j The closer to 1, the more accurate the prediction, the smaller the loss value.
The experiment adopts a Deeplab V3+ network model (hereinafter, Deeplab V3+ _ MobileNet V2) based on the optimization of a MobileNet V2 network, and the overall architecture is shown in FIG. 5. The method has the advantages that the MobileNetV2 is used for replacing an Xception network as a main feature extraction network, a Deeplab V3+ algorithm is optimized, the number of model parameters can be greatly reduced, the lightweight design of a model is realized, feature extraction of deep convolution can be guaranteed to be completed at a high dimension, the model calculation performance is improved, and the network structure is shown in figure 5.
In an encoder, extracting hot spot features by utilizing a MobileNet V2 network and cavity convolutions with the rates of 6, 12 and 18 respectively, merging, performing 1 × 1 convolution compression, and outputting high-level features; in the decoder, the low-level features in the encoder are subjected to 1 × 1 convolution dimensionality reduction, then are subjected to feature fusion with the high-level features, the target boundary is restored, and then a final prediction graph is decoded through 3 × 3 convolution and 4 times upsampling.
Down-sampling of input images is required in an encoderThe down-sampling functions not only to fit an image to the size of a display area but also to generate a thumbnail of a corresponding image, and for example, a 512 × 512 input image is down-sampled by 16 times to obtain a result
An image of a resolution of size. The larger the down-sampling multiple is, the smaller the output image is, so the 16-time down-sampling is more suitable for the segmentation with larger targets, and the small target error of the hot spot of the photovoltaic panel is very large, so the down-sampling multiple is changed from 16 to 8, and the segmentation effect is improved.
In the aspect of hot spot segmentation of a photovoltaic panel, because the hot spots are small, the foreground and background are extremely unbalanced, and a common cross entropy Loss function cannot solve the problem of extreme unbalance, the method provided by the invention improves the segmentation precision by combining a sieve Loss (Dice Loss) function and a cross entropy Loss function, wherein the Dice Loss can be expressed as formula (6):
h is the height of the image, W is the width of the image, p i，j Is the predicted probability, g, of having a foreground at position (i, j) i，j Is the label (0 or 1) at position (i, j).
And the target loss caused by downsampling is improved, and the segmentation precision is further improved by modifying the loss function into a mode of combining a Dice loss function and a cross entropy loss function.
Another object of the present embodiment is to provide a photovoltaic panel hot spot detection apparatus based on a deep convolutional neural network, which includes an image acquisition unit, a recognition unit, and a segmentation unit.
The image acquisition unit is used for acquiring an infrared image of the photovoltaic panel; the identification unit is used for constructing a photovoltaic panel identification model by utilizing an improved Yolov4 algorithm, identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model; the method comprises the following steps that a segmentation unit utilizes an optimized semantic segmentation algorithm DeepLabV3+ to construct a hot spot segmentation model, and hot spots on a intercepted photovoltaic panel are detected and segmented through the hot spot segmentation model; the improvement of the target detection algorithm Yolov4 comprises replacing the main feature extraction network CSPDarknet53 of the Yolov4 algorithm with a lightweight network MobileNet V2, and replacing the standard 3 × 3 convolution in the enhanced feature extraction network PANet with a deep separable convolution; the optimization of the semantic segmentation algorithm DeepLabV3+ includes replacing the backbone feature extraction network Xconcept with a lightweight network MobileNetV 2.
The method provided by the invention is adopted to detect the hot spots of the photovoltaic panel, and corresponding detection results and analysis results are given.
First, test set-up
Training and testing platform
The test processing platform is an Inter (R) Xeon (R) Bronze 3204CPU, a main frequency 1.90GHz and 32GB memory, a video card is a NVIDA Quadro P5000 workstation, an operating system is windows1064 bit, and the environment is set up by using a pytorch frame.
Transfer learning
Because the number of data sets is small, the convergence effect of the direct training model is possibly poor, and the high recognition rate is not achieved. Therefore, the trained model on a complete data set can be applied to photovoltaic panel recognition and hot spot segmentation through transfer learning, so that the model can be rapidly converged under the condition of a smaller data set, and higher accuracy is realized.
Training parameters
(1) Photovoltaic panel recognition model parameter training
And (3) expanding 6228 photovoltaic panel infrared images after data expansion according to the proportion of a training set to a verification set of 8: 2 for carrying out the training of the Yolov4 model. The training process is divided into two stages, namely a freezing stage and a thawing stage, wherein the trunk network is frozen in the freezing stage, the feature extraction network is not changed, the trunk is not frozen in the thawing stage, and the feature extraction network is changed. And using the trained pre-training weight, wherein the initial learning rate is 1e-3 in the freezing stage and is trained for 50 times, and the initial learning rate is 1e-4 in the thawing stage and is trained for 50 times, and the training times are 100 times in total. After the backbone network was replaced, each model was tested 3 times and the average was taken as the test result.
(2) Hot spot segmentation model parameter training
Screening out pictures 1610 with hot spots in the photovoltaic panel identification result pictures according to the proportion of a training set to a verification set of 9: 1 for training the deplab v3+ model. And (3) using the trained pre-training weights, wherein the initial learning rate is 1e-3 in the freezing stage and is trained for 35 times, and the initial learning rate is 1e-4 in the thawing stage and is trained for 35 times, and the training times are 70 times in total. The results were also obtained by averaging the respective indices in 3 trials.
Second, photovoltaic panel test results and analysis
Criteria for discrimination
The test set the intersection ratio (IOU) threshold of the predicted result to the actual target to 0.5, which is the correct prediction when the IOU value is greater than 0.5. The prediction effects of different algorithms are measured by accuracy (Precision), average detection Precision (AP) and recall (recall) respectively. The feasibility of the practical application of the model is measured by using an FPS (field programmable gate array) and a model size, wherein the FPS is the number of frames per second of a detected picture, the model size is determined by model parameters, and the indexes are defined as (7) to (10):
in the above equation, tp represents the number of actually positive samples and predicted positive samples, fp represents the number of actually positive samples and predicted negative samples, fn represents the number of actually negative samples and predicted positive samples, s represents the number of processed pictures, and t represents the time required to process a picture.
Analysis of network effects on model Performance
In order to verify the validity of the detection model, the Yolov4 model replacing the backbone network is tested, the performance of the MobileNetV2-Yolov4-lite network model provided by the invention is tested by comparing the AP, FPS, Precision, parameter and model size of each algorithm, and the test results are shown in table 1:
table 1: comparison of effects of different detection models on photovoltaic panel infrared image dataset
As can be seen from Table 1, different models perform better on AP, the AP index of the traditional Yolov4 algorithm is relatively higher and is 99.66%, which is 0.1% higher than that of the algorithm of the invention, but the Yolov4 model has a size which is 197.6M larger than that of the V2 of the invention, and the parameter quantity is 5 times that of the invention, so that the effect of replacing the network in practical application is obvious. On recall and precision, the difference of models is not large, and the recognition effect meets the target detection requirement, but on FPS, the V2 is the highest and is 22.1 frames/second, 8.4 frames/second higher than Yolov4, 3.9 frames/second higher than V1, and 6.2 frames/second higher than V3. In terms of model size, the V2 of the invention is the smallest, 46.4M, 7.2M smaller than V3 and 4.6M smaller than V1. Therefore, the MobileNet V2-Yolov4-lite model has great advantages in detection speed and calculation parameters while ensuring precision, and can be suitable for real-time detection and the condition of low hardware resources.
The photovoltaic panel infrared image dataset was tested under the MobileNetV2-Yolov4-lite model of the present invention as shown in fig. 8 and 9.
Thirdly, photovoltaic panel hot spot segmentation result and analysis
Criteria for discrimination
The invention adopts average pixel accuracy (MPA) and average cross-over ratio (MIoU) to measure the segmentation effect, and FPS and model size are used to analyze whether practical application is feasible. Each index is defined as formula (11), (12):
in the formula: pij represents the number of pixels with real value i but predicted to be j; n represents the number of test classes.
Test results and analysis
Testing the influence of different modification methods on the model, taking A as a basic model, respectively replacing a Deeplab V3+ main network with a MobileNet V2, modifying down-sampling 16 to 8, modifying a Loss function to be a Dice Loss function and combining the Dice Loss function with a cross entropy Loss function, and carrying out training and testing tests, wherein the modification mode is shown in Table 2:
table 2: models of different processing methods
In order to verify the validity of the detection model of the invention, 5 models of different processing methods in the table above are respectively tested, the performance of the network model provided by the invention is detected by comparing the MPA, MIoU, FPS and model size of each algorithm, and the test result is shown in table 3:
TABLE 3 Effect of different methods on the model
As shown in table 3, the replacement backbone network is MobileNetV2, the B model is greatly improved in each index compared with a model a, the MPA is improved by 18.77%, the MIoU is improved by 9.36%, the FPS is improved by nearly 1 time, and the model is reduced by 186.7M, which indicates that the effect of replacing the backbone network is very good. After the downsampling modification is 8, although the C-to-B model is dropped by 2.5 frames/second on the FPS, there are improvements on MPA and MIoU, which are 2.92% and 3.58%, respectively, and it can be seen that the modification of the downsampling multiple improves the model segmentation capability. After the loss function is modified into two combinations, the D is improved compared with the B model on each index, the FPS is improved by 0.2 frame/second, the MPA is improved by 1.18 percent, and the MIoU is improved by 0.93 percent, so that the model segmentation effect is obviously improved by modifying the loss function. The model E not only modifies the downsampling multiple, but also modifies the loss function, has great advantages in prediction precision and model size under comprehensive comparison, and can be suitable for the conditions of real-time detection and low hardware resources.
In the present invention, under the condition that the deep sampling and loss function model is modified by the Deeplab V3+ _ MobileNetV2, the infrared image dataset of the photovoltaic panel is tested as shown in FIGS. 10 and 11, wherein (a) is a set of original images before being segmented, (b) is a set of images with background after being segmented, and (c) is a set of images with background after being segmented.
The photovoltaic panel hot spot identification model of Yolov4+ MobileNet V2+ Deeplabv3+ researched and designed by the invention can effectively realize the target identification and hot spot positioning of the photovoltaic panel, and has the following advantages:
(1) the infrared image photovoltaic panel identification model is based on an improved Yolov4 algorithm, tests a photovoltaic panel infrared image data set subjected to data amplification, comprehensively considers calculation speed and parameter quantity, and can meet the identification of photovoltaic panels in rural complex backgrounds through a preferred Yolov4+ MobileNet V2 model.
(2) The infrared image hot spot segmentation model replaces a main network with a MobileNet V2 model, the down-sampling multiple is changed into 8 model, and a Deeplab V3+ _ MobileNet V2 model combining a Dice loss function and a cross entropy loss function is used.
(3) The hot spot detection method for the photovoltaic panel can accurately identify the hot spot of the photovoltaic panel, provides theoretical support for the hot spot detection of the aerial photo photovoltaic panel, and provides technical support for rural distributed photovoltaic panel detection.
Test results show that the method can accurately identify the hot spots of the photovoltaic panel, the identification accuracy rate of the photovoltaic panel is 99.56%, and the detection speed is 22.1 frames/second. The hot spot segmentation accuracy after the photovoltaic panel is identified reaches 95.99%, the intersection ratio mIou reaches 85.58, and the detection speed is 24.5 frames/second. The photovoltaic panel hot spot detection method provided by the invention can meet the requirement of rural photovoltaic panel fault detection and provide technical support for the rapid development of future rural photovoltaic poverty-relieving power stations.
The above-mentioned embodiments are only preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, and any simple modifications or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.
Claims (9)
1. A photovoltaic panel hot spot detection method based on a deep convolutional neural network is characterized by comprising the following steps:
acquiring an infrared image of the photovoltaic panel;
constructing a photovoltaic panel identification model by using an improved target detection algorithm Yolov4, and identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model;
constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting hot spots on the intercepted photovoltaic panel through the hot spot segmentation model;
the improvement of the target detection algorithm Yolov4 comprises replacing the main feature extraction network CSPDarknet53 of the Yolov4 algorithm with a lightweight network MobileNet V2, and replacing the standard 3 × 3 convolution in the enhanced feature extraction network PANet with a deep separable convolution; the optimization of the semantic segmentation algorithm DeepLabV3+ includes replacing the backbone feature extraction network Xconcept with a lightweight network MobileNetV 2.
2. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 1, wherein the photovoltaic panel infrared image is obtained by unmanned aerial vehicle aerial photography.
3. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 1, wherein the photovoltaic panel in the photovoltaic panel infrared image is identified and intercepted by the photovoltaic panel identification model, comprising the following steps:
performing primary feature extraction on the input photovoltaic panel infrared image through a main feature extraction network MobileNetV2 of the improved target detection algorithm Yolov4 to obtain a primary feature layer;
performing pooling processing of different scales on the feature layer through a spatial pyramid pooling network (SPP) of the improved target detection algorithm Yolov4, and performing 3 × 3 convolution after cascading;
inputting the convolved features into an enhanced feature extraction network (PAne) for feature fusion to obtain four coordinates top, left, bottom and right of a prediction frame;
and identifying the photovoltaic panel identification model by using a matrix form through the four coordinates top, left, bottom and right of the obtained prediction frame to obtain a photovoltaic panel identification result picture and intercepting the photovoltaic panel identification result picture.
4. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 1, wherein the hot spot on the intercepted photovoltaic panel is detected and segmented by the hot spot segmentation model, and the method comprises the following steps:
in an encoder of the optimized semantic segmentation algorithm DeepLabV3+, extracting hot spot features on a photovoltaic panel by utilizing a cavity convolution of a MobileNet V2 network, merging the hot spot features, performing 1 × 1 convolution compression, and outputting advanced features;
in a decoder of the optimized semantic segmentation algorithm DeepLabV3+, the low-level features in the encoder are subjected to 1 × 1 convolution dimensionality reduction, then are subjected to feature fusion with the high-level features, the target boundary is restored, and the hot spots on the final photovoltaic panel are decoded.
5. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 4, wherein the loss function adopted by the optimized semantic segmentation algorithm DeepLabV3+ is a loss function obtained by combining a Dice loss function and a cross entropy loss function, and the expression is as follows:
wherein H is the height of the image, W is the width of the image, p i，j Is a predicted probability, g, of the foreground at position (i, j) i，j The label at position (i, j) is 0 or 1.
6. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 2, wherein the photovoltaic panel infrared image is preprocessed by a data augmentation technology before the photovoltaic panel identification is performed by the photovoltaic panel identification model, so that data augmentation is realized.
7. The photovoltaic panel hot spot detection method based on the deep convolutional neural network as claimed in claim 6, wherein the photovoltaic panel infrared image data is preprocessed by using a data augmentation technology, and the method comprises the following steps:
randomly rotating the photovoltaic panel infrared image by 0-120 degrees to form an expanded photovoltaic panel infrared image;
or translating the photovoltaic panel infrared image to the right and rotating the photovoltaic panel infrared image by 180 degrees to form an expanded photovoltaic panel infrared image so that the detection target is positioned at different positions;
or the infrared image of the photovoltaic panel is darkened to form an expanded infrared image of the photovoltaic panel, so that the problem occurring in actual shooting is simulated;
and marking the expanded photovoltaic panel infrared image for network training.
8. The photovoltaic panel hot spot detection method based on the deep convolutional neural network of claim 7, further comprising performing parameter training on the photovoltaic panel recognition model and the hot spot segmentation model by using the extended photovoltaic panel infrared image:
the parameter training of the photovoltaic panel recognition model comprises:
and (3) the expanded photovoltaic panel infrared image is processed according to the proportion of a training set to a verification set of 8: 2, dividing the training process into a freezing stage and a thawing stage;
in the freezing stage, the main network is frozen, and the feature extraction network is not changed; in the unfreezing stage, the trunk is not frozen, and the feature extraction network is changed;
using the trained pre-training weight, wherein the initial learning rate is 1e-3 in the freezing stage and is trained for 50 times, the initial learning rate is 1e-4 in the thawing stage and is trained for 50 times, and the training times are 100 times in total; after the trunk network is replaced, each model is subjected to repeated tests for 3 times, and the average value is taken as a test result;
the parameter training of the hot spot segmentation model comprises:
screening out a plurality of pictures with hot spots in the pictures of the identification results of the photovoltaic panel according to the proportion of a training set to a verification set 9: 1, dividing;
using the trained pre-training weight, wherein the initial learning rate is 1e-3 in the freezing stage and is trained for 35 times, and the initial learning rate is 1e-4 in the thawing stage and is trained for 35 times, and the training times are 70 times; the results were also obtained by averaging the respective indices in 3 trials.
9. A photovoltaic panel hot spot detection device based on a deep convolutional neural network is characterized by comprising the following components:
the image acquisition unit is used for acquiring an infrared image of the photovoltaic panel;
the identification unit is used for constructing a photovoltaic panel identification model by utilizing an improved Yolov4 algorithm, identifying and intercepting the photovoltaic panel in the infrared image of the photovoltaic panel through the photovoltaic panel identification model;
the segmentation unit is used for constructing a hot spot segmentation model by using an optimized semantic segmentation algorithm DeepLabV3+, and detecting and segmenting hot spots on the intercepted photovoltaic panel through the hot spot segmentation model;
the improvement of the target detection algorithm Yolov4 comprises replacing the main feature extraction network CSPDarknet53 of the Yolov4 algorithm with a lightweight network MobileNet V2, and replacing the standard 3 × 3 convolution in the enhanced feature extraction network PANet with a deep separable convolution; the optimization of the semantic segmentation algorithm DeepLabV3+ includes replacing the backbone feature extraction network Xconcept with a lightweight network MobileNetV 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210587259.5A CN114973032B (en) | 2022-05-27 | 2022-05-27 | Deep convolutional neural network-based photovoltaic panel hot spot detection method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210587259.5A CN114973032B (en) | 2022-05-27 | 2022-05-27 | Deep convolutional neural network-based photovoltaic panel hot spot detection method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114973032A true CN114973032A (en) | 2022-08-30 |
| CN114973032B CN114973032B (en) | 2024-06-18 |
Family
ID=82954772
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210587259.5A Active CN114973032B (en) | 2022-05-27 | 2022-05-27 | Deep convolutional neural network-based photovoltaic panel hot spot detection method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114973032B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115641434A (en) * | 2022-12-26 | 2023-01-24 | 浙江天铂云科光电股份有限公司 | Power equipment positioning method, system, terminal and storage medium |
| CN116580363A (en) * | 2023-07-14 | 2023-08-11 | 尚特杰电力科技有限公司 | Photovoltaic panel hot spot identification method, storage medium and electronic equipment |
| CN116912162A (en) * | 2023-04-25 | 2023-10-20 | 东莞市华视智能科技有限公司 | Tobacco leaf mottle extraction method based on computer vision deep learning |
| CN117129088A (en) * | 2023-08-08 | 2023-11-28 | 上海根根信息技术有限公司 | Chip temperature testing method and system |
| CN117689271A (en) * | 2024-02-02 | 2024-03-12 | 深圳市崇禧精密金属制品有限公司 | Quality management method and device for product, terminal equipment and storage medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112164038A (en) * | 2020-09-16 | 2021-01-01 | 上海电力大学 | Photovoltaic hot spot detection method based on deep convolutional neural network |
| CN112200764A (en) * | 2020-09-02 | 2021-01-08 | 重庆邮电大学 | Photovoltaic power station hot spot detection and positioning method based on thermal infrared image |
| CN112508977A (en) * | 2020-12-29 | 2021-03-16 | 天津科技大学 | Deep learning-based semantic segmentation method for automatic driving scene |
| CN112634261A (en) * | 2020-12-30 | 2021-04-09 | 上海交通大学医学院附属瑞金医院 | Stomach cancer focus detection method and device based on convolutional neural network |
| WO2021244079A1 (en) * | 2020-06-02 | 2021-12-09 | 苏州科技大学 | Method for detecting image target in smart home environment |
| CN113837994A (en) * | 2021-07-29 | 2021-12-24 | 尚特杰电力科技有限公司 | Photovoltaic panel defect diagnosis method based on edge detection convolutional neural network |
| CN114299033A (en) * | 2021-12-29 | 2022-04-08 | 中国科学技术大学 | YOLOv 5-based photovoltaic panel infrared image hot spot detection method and system |
| US20220129685A1 (en) * | 2020-10-23 | 2022-04-28 | Deka Products Limited Partnership | System and Method for Determining Object Characteristics in Real-time |
-
2022
- 2022-05-27 CN CN202210587259.5A patent/CN114973032B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021244079A1 (en) * | 2020-06-02 | 2021-12-09 | 苏州科技大学 | Method for detecting image target in smart home environment |
| CN112200764A (en) * | 2020-09-02 | 2021-01-08 | 重庆邮电大学 | Photovoltaic power station hot spot detection and positioning method based on thermal infrared image |
| CN112164038A (en) * | 2020-09-16 | 2021-01-01 | 上海电力大学 | Photovoltaic hot spot detection method based on deep convolutional neural network |
| US20220129685A1 (en) * | 2020-10-23 | 2022-04-28 | Deka Products Limited Partnership | System and Method for Determining Object Characteristics in Real-time |
| CN112508977A (en) * | 2020-12-29 | 2021-03-16 | 天津科技大学 | Deep learning-based semantic segmentation method for automatic driving scene |
| CN112634261A (en) * | 2020-12-30 | 2021-04-09 | 上海交通大学医学院附属瑞金医院 | Stomach cancer focus detection method and device based on convolutional neural network |
| CN113837994A (en) * | 2021-07-29 | 2021-12-24 | 尚特杰电力科技有限公司 | Photovoltaic panel defect diagnosis method based on edge detection convolutional neural network |
| CN114299033A (en) * | 2021-12-29 | 2022-04-08 | 中国科学技术大学 | YOLOv 5-based photovoltaic panel infrared image hot spot detection method and system |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115641434A (en) * | 2022-12-26 | 2023-01-24 | 浙江天铂云科光电股份有限公司 | Power equipment positioning method, system, terminal and storage medium |
| CN116912162A (en) * | 2023-04-25 | 2023-10-20 | 东莞市华视智能科技有限公司 | Tobacco leaf mottle extraction method based on computer vision deep learning |
| CN116580363A (en) * | 2023-07-14 | 2023-08-11 | 尚特杰电力科技有限公司 | Photovoltaic panel hot spot identification method, storage medium and electronic equipment |
| CN116580363B (en) * | 2023-07-14 | 2023-09-26 | 尚特杰电力科技有限公司 | Photovoltaic panel hot spot identification method, storage medium and electronic equipment |
| CN117129088A (en) * | 2023-08-08 | 2023-11-28 | 上海根根信息技术有限公司 | Chip temperature testing method and system |
| CN117689271A (en) * | 2024-02-02 | 2024-03-12 | 深圳市崇禧精密金属制品有限公司 | Quality management method and device for product, terminal equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114973032B (en) | 2024-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114973032A (en) | Photovoltaic panel hot spot detection method and device based on deep convolutional neural network | |
| CN111784633B (en) | Insulator defect automatic detection algorithm for electric power inspection video | |
| CN113065558A (en) | Lightweight small target detection method combined with attention mechanism | |
| CN110610475B (en) | Visual defect detection method of deep convolutional neural network | |
| CN114022420A (en) | Detection method for automatically identifying defects of photovoltaic cell EL assembly | |
| CN111680705B (en) | MB-SSD method and MB-SSD feature extraction network suitable for target detection | |
| US20220366682A1 (en) | Computer-implemented arrangements for processing image having article of interest | |
| CN114155474A (en) | Damage identification technology based on video semantic segmentation algorithm | |
| NL2030315B1 (en) | Computer-implemented ai method for detecting surface defects of electroluminescent semiconductor panel | |
| CN114972312A (en) | Improved insulator defect detection method based on YOLOv4-Tiny | |
| CN116503318A (en) | Aerial insulator multi-defect detection method, system and equipment integrating CAT-BiFPN and attention mechanism | |
| CN115830004A (en) | Surface defect detection method, device, computer equipment and storage medium | |
| CN114283137A (en) | Photovoltaic module hot spot defect detection method based on multi-scale characteristic diagram inference network | |
| CN115908354A (en) | Photovoltaic panel defect detection method based on double-scale strategy and improved YOLOV5 network | |
| CN116612106A (en) | Method for detecting surface defects of optical element based on YOLOX algorithm | |
| CN114596278A (en) | Method and device for detecting hot spot defects of photovoltaic panel of photovoltaic power station | |
| CN116258664A (en) | Deep learning-based intelligent defect detection method for photovoltaic cell | |
| CN116385911A (en) | Lightweight target detection method for unmanned aerial vehicle inspection insulator | |
| CN109409285B (en) | Remote sensing video target detection method based on overlapped slices | |
| CN111161156A (en) | Deep learning-based underwater pier disease image resolution enhancement method | |
| CN114549489A (en) | Carved lipstick quality inspection-oriented instance segmentation defect detection method | |
| CN116485802B (en) | Insulator flashover defect detection method, device, equipment and storage medium | |
| CN116958073A (en) | Small sample steel defect detection method based on attention feature pyramid mechanism | |
| CN116433591A (en) | Method for detecting hot spots of photovoltaic panel and calculating area ratio | |
| CN116883893A (en) | Tunnel face underground water intelligent identification method and system based on infrared thermal imaging |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |