CN116052082A - Power distribution station room anomaly detection method and device based on deep learning algorithm - Google Patents
Power distribution station room anomaly detection method and device based on deep learning algorithm Download PDFInfo
- Publication number
- CN116052082A CN116052082A CN202310049195.8A CN202310049195A CN116052082A CN 116052082 A CN116052082 A CN 116052082A CN 202310049195 A CN202310049195 A CN 202310049195A CN 116052082 A CN116052082 A CN 116052082A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- deep learning
- abnormal
- 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.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 134
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 38
- 238000013135 deep learning Methods 0.000 title claims abstract description 38
- 238000009826 distribution Methods 0.000 title claims abstract description 32
- 230000002159 abnormal effect Effects 0.000 claims abstract description 63
- 238000000034 method Methods 0.000 claims abstract description 41
- 230000007246 mechanism Effects 0.000 claims abstract description 18
- 238000007781 pre-processing Methods 0.000 claims abstract description 10
- 230000002776 aggregation Effects 0.000 claims abstract description 7
- 238000004220 aggregation Methods 0.000 claims abstract description 7
- 238000012549 training Methods 0.000 claims description 41
- 238000012545 processing Methods 0.000 claims description 27
- 230000006870 function Effects 0.000 claims description 12
- 238000003860 storage Methods 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 10
- 230000000694 effects Effects 0.000 claims description 9
- 238000013138 pruning Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 230000001629 suppression Effects 0.000 claims description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 7
- 230000005856 abnormality Effects 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 238000005520 cutting process Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 230000005764 inhibitory process Effects 0.000 claims description 4
- 238000013519 translation Methods 0.000 claims description 4
- 230000004069 differentiation Effects 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000004044 response Effects 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims 1
- 238000005516 engineering process Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 230000015654 memory Effects 0.000 description 7
- 241001465754 Metazoa Species 0.000 description 6
- 230000006399 behavior Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 201000010099 disease Diseases 0.000 description 3
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 3
- 230000009545 invasion Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000012876 topography Methods 0.000 description 3
- 240000006365 Vitis vinifera Species 0.000 description 2
- 235000014787 Vitis vinifera Nutrition 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000011897 real-time detection Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 206010000117 Abnormal behaviour Diseases 0.000 description 1
- 241000282472 Canis lupus familiaris Species 0.000 description 1
- 241000282326 Felis catus Species 0.000 description 1
- 244000141353 Prunus domestica Species 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 235000019580 granularity Nutrition 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000033001 locomotion Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- LLLILZLFKGJCCV-UHFFFAOYSA-M n-methyl-n-[(1-methylpyridin-1-ium-4-yl)methylideneamino]aniline;methyl sulfate Chemical compound COS([O-])(=O)=O.C=1C=CC=CC=1N(C)\N=C\C1=CC=[N+](C)C=C1 LLLILZLFKGJCCV-UHFFFAOYSA-M 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000012360 testing method 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
- G06V20/50—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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- 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
- G06V10/24—Aligning, centring, orientation detection or correction of the image
- G06V10/245—Aligning, centring, orientation detection or correction of the image by locating a pattern; Special marks for positioning
-
- 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
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- 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
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
-
- 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
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Abstract
The invention discloses a power distribution station room anomaly detection method and device based on a deep learning algorithm, wherein the method comprises the following steps: acquiring an image in a power distribution station room, and preprocessing the image, wherein the preprocessing comprises image scaling and image patching; inputting the preprocessed image into a trained abnormal target detection model based on a deep learning algorithm; and determining the abnormal detection result of the power distribution station room according to the output of the abnormal target detection model. The abnormal target detection model comprises an input layer, a main network with a CA attention mechanism, a neck network and an output layer; the backbone network is formed by combining a series of convolution layers, and a CA attention mechanism is used for fusing position information; the neck network adopts the structure of a feature map pyramid network FPN+pixel aggregation network PAN, and the FPN is combined with the PAN to obtain a feature map which is finally predicted.
Description
Technical Field
The invention belongs to the technical field of abnormal detection of power distribution station rooms, and relates to a power distribution station room abnormal detection method and device based on a deep learning algorithm.
Background
With the rapid development of the Internet industry, the artificial intelligence technology tends to be mature, and the power distribution station room of the power industry gradually tends to be unmanned and intelligent, so that the waste of human resources is effectively relieved while the power operation efficiency is improved, and great economic benefits and social benefits are created. However, in the unattended power distribution station room, the frequency of the occurrence of violations or abnormal conditions is increased rapidly, personnel do not wear safety helmets to violate the entering station room, personnel fall down to endanger life safety, the problem that small animals in the power distribution station room invade the room to damage electric power facilities and damage lines to cause safety accidents and the like occurs, the safety environment in the station room is greatly influenced, a large amount of casualties and huge economic losses are caused, and the intelligent power grid work propulsion is limited.
The current prevention means for the biological invasion problem is to prevent the small animals from entering the room by using the traditional modes such as wire netting, baffle plates, mouse sticking plates and the like, and the modes can have certain defense effect, but have certain problems at the same time. First, the defenses of these methods are limited and animals can enter the station house by other routes; secondly, devices such as baffles are easily damaged or aged and corroded by animals, and the devices need to be regularly overhauled and replaced, so that the traditional method cannot well meet the requirements. And to personnel fall down and not wear the condition of safety helmet, can only detect through the control by the manual work generally, inefficiency and cause the waste of manpower resources.
Disclosure of Invention
The purpose is as follows: in order to overcome the defects in the prior art, the invention provides a substation room anomaly detection method and device based on a deep learning algorithm, so as to realize automatic detection of biological invasion conditions in unmanned substation rooms and realize detection of falling and illegal behaviors of personnel.
Deep learning is a new research direction in the field of machine learning, and it was introduced into machine learning to make it closer to the original goal-artificial intelligence.
Deep learning is the inherent regularity and presentation hierarchy of learning sample data, and the information obtained during such learning is helpful in interpreting data such as text, images and sounds. Its final goal is to have the machine have analytical learning capabilities like a person, and to recognize text, image, and sound data. Deep learning is a complex machine learning algorithm that achieves far greater results in terms of speech and image recognition than prior art. Target detection, also called target extraction, is an image segmentation based on target geometry and statistical features. The method combines the segmentation and the identification of the target into a whole, and the accuracy and the real-time performance are an important capability of the whole system.
At present, the detection precision of the target detection method based on deep learning is continuously increased, so that the technology is continuously applied to various fields. In the economic field, gray Relational Analysis (GRA) and artificial neural network models are utilized for predicting consumer exchange traded funds or stock trends using deep neural networks. In the industrial field, the best known and popular automatic driving of the internet of vehicles is to detect targets such as lane lines, traffic lights and the like by using a deep learning technology and feed back abnormal information around the vehicles. In addition, on the industrial assembly line, defect detection is one of main research directions of current target detection, unqualified objects are screened out by utilizing the difference between defective products and qualified products, and quality detection efficiency is greatly improved. The target detection technology is also used for amplifying the wonderful colors in the agricultural field, and a novel deep learning system structure VddNet (grape vine disease detection network) is provided for detecting grape vine diseases, and the knn algorithm is used for detecting citrus yellow dragon diseases. Zeze et al have implemented apple recognition using CNN.
In summary, the image is subjected to target detection by using a deep learning technology, so that the positioning, recognition and feedback of the target in the image can be realized, and the powerful detection capability is provided for the abnormality detection method provided by the invention, thereby achieving the purpose of checking and preventing the abnormal situation and the illegal behavior in the station building. In the scheme, the station room intrusion detection method based on deep learning is provided, and the scheme aims at the situation in the power distribution station room, improves a deep learning model and realizes real-time anomaly detection in the station room. Specifically, a lightweight anomaly detection model is constructed by utilizing pruning technology and attention mechanism technology. The scheme is based on the following points: (1) Adding an attention mechanism to construct a model with stronger generalization capability, and improving the object detection capability of weaker light or blurring; (2) Pruning is carried out on the model, so that a model with smaller model volume and faster recognition speed is obtained, and the model can be deployed in an embedded terminal for detection. The model can be used for detecting violations, anomalies and intrusion problems in the substation room in real time after the camera shoots the image.
The technical scheme is as follows: in order to solve the technical problems, the invention adopts the following technical scheme:
in a first aspect, a substation room anomaly detection method based on a deep learning algorithm is provided, including:
acquiring an image in a power distribution station room, and preprocessing the image, wherein the preprocessing comprises image scaling and image patching;
inputting the preprocessed image into a trained abnormal target detection model based on a deep learning algorithm;
determining an abnormal detection result of the power distribution station room according to the output of the abnormal target detection model;
the abnormal target detection model comprises an input layer, a main network with a CA attention mechanism, a neck network and an output layer; the backbone network is formed by combining a series of convolution layers, and a CA attention mechanism is used for fusing position information; the neck network adopts a structure of a feature map pyramid network FPN+ pixel aggregation network PAN, wherein the feature map pyramid network FPN is of a top-down network structure, high-level feature maps are transferred and fused through up-sampling operation, high-level strong semantic features are conveyed, convolution operation is carried out on the pixel aggregation network PAN from bottom to top, strong positioning features are transferred, and the detection capability of targets with different scales is improved; and combining the FPN with the PAN to obtain a final predicted feature map.
In some embodiments, the processing of the abnormal target detection model includes:
the image is sent into a backbone network through an input layer, and the feature extraction is carried out on the input image through the backbone network to generate a high-level feature map;
the neck network performs differentiation detection on the high-level feature images, performs cross-layer splicing through the up-sampling network and the convolutional neural network, and performs target detection on the feature images which are divided into different sizes;
and classifying the feature map output by the neck network through a convolution layer of the output layer to obtain an abnormal target detection result.
Further, in some embodiments, the division into different size feature maps performs different size object detection, including: a feature map for detecting the size of small targets 76 x 76, a feature map for detecting the size of medium targets 38 x 38 and a feature map for detecting the size of large targets 19 x 19 are obtained, respectively.
In some embodiments, the training method of the abnormal target detection model includes:
acquiring a training data set;
and inputting the training data set into an abnormal target detection model to be trained for training, carrying out weighted non-maximum suppression processing and confidence assessment on the output of the model, and carrying out optimization updating on model parameters by combining a loss function until a preset condition is reached, so as to obtain the trained abnormal target detection model.
Further, acquiring the training data set further comprises: and carrying out data enhancement processing on the training data set, wherein the data enhancement processing comprises Mosaic data enhancement, image overlapping and dropout noise addition processing.
Further, the mosaics are spliced by adopting 4 pictures in a random zooming, random cutting and random arrangement mode;
the image overlapping is to fuse two pictures with the same size according to random proportion to synthesize a new picture, and the proportion numerical values obey Beta distribution with parameters; the new picture is provided with the targets of the two pictures, and the labels of the two pictures are spliced together to form a new label;
the dropout noise is added to randomly generate a plurality of rectangular blocks with the same size on the image, so that the original picture is shielded, and the information of the positions of the rectangular blocks is lost.
Further, inputting the training data set into an abnormal target detection model to be trained for training, and carrying out weighted non-maximum suppression processing and confidence assessment on the output of the model, wherein the method comprises the following steps:
the output of the abnormal target detection model comprises the score, the confidence score and the position information of each category of the detected candidate target frame;
performing non-maximum value inhibition processing on each candidate target frame, and removing redundant target frames to obtain a high-quality target frame;
and comparing the confidence coefficient of the high-quality target frame with a preset confidence coefficient threshold value, judging whether the target frame is qualified or not in response to the fact that the confidence coefficient is larger than the preset confidence coefficient threshold value, storing the related target frame and the picture, and returning a result.
Further, the training method of the abnormal target detection model further comprises pruning the abnormal target detection model:
constraint is carried out by adding an L1 penalty term to a BN layer after a convolution layer in the neural network, so that model parameters are sparse; the output OUT formula of BN layer is as follows:
OUT＝γ*x i ′ +β
wherein, gamma is a scaling factor, and beta is a translation factor; x is x i ′ The input of the BN layer;
the gamma is normally distributed, a large number of gamma values tend to 0 by adding an L1 penalty term, and more layers which do not influence the result are removed;
wherein L is a loss function of model training, the first term in the formula (x, l (f (x, W, y) is the loss of model training, x is the input data, and the actual prediction result is calculated by the model weight W, the result f (x, W and the label result y calculate loss of loss, the second term lambda sigma) γ∈r gγ) is a constraint condition, where γ is a polar radius, gγ) is a sum of absolute values of elements in a weight vector, λ is an adjustable regularization coefficient, and by setting a proper λ value, the weight achieves a sparsifying effect.
In a second aspect, the invention provides a substation room anomaly detection device based on a deep learning algorithm, which comprises a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is configured to operate in accordance with the instructions to perform the steps of the method according to the first aspect.
In a third aspect, the present invention provides a storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method of the first aspect.
In a fourth aspect, the present invention provides a computing device comprising:
one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods of the first aspect.
The method provided by the invention adopts a one-stage target detection mechanism for intrusion detection, and relies on a YOLOv5 target detection model to identify biological intrusion, personnel fall and illegal behaviors in a substation room. The method is selected to construct an abnormality detection mechanism based on the following reasons that (1) the YOLOv5 algorithm model has good effect in terms of detection speed, and an abnormality detection system formed by improving the YOLOv5 algorithm model can rapidly scan and detect object detection in a power distribution station room, so that the detection speed is improved, and the effect of real-time detection is achieved. (2) The detection result can be displayed and stored more intuitively, and the detection result is convenient for a user to observe and take. (3) The system is convenient to deploy and floor, and can be directly deployed in different devices by simple conversion and modification.
The beneficial effects are that: the power distribution station room anomaly detection method and device based on the deep learning algorithm provided by the invention have the following advantages: the detection and the positioning in the system are completed for the abnormal conditions in the power distribution station room by utilizing a convolutional neural network and a computer vision technology, and a lightweight and strong interference abnormal detection model is constructed, so that the volume and the generalization capability of the model are effectively reduced on the premise of not reducing the detection precision, and the real-time monitoring in the station room and the recognition of abnormal and illegal conditions are completed. The following is a detailed description.
High efficiency: many current detection studies in the distribution station room focus on strengthening and improving by physical measures and by infrared sensing and other devices, and even manual detection. The existing target detection algorithm is directly applied to a complex station room environment, and has high false alarm rate and low efficiency. The abnormality detection method provided by the invention provides an abnormality detection model based on deep learning and model pruning technology under the condition of not reducing detection precision, and the system prunes through a one-stage target detection algorithm, so that the detection speed is high, the deployment condition is low, the portability is strong, and the smooth and efficient detection of abnormal conditions in a station room can be realized.
The universality is strong: in conventional solutions, various disadvantages need to be overcome due to topography factors and operation and maintenance tests. For example, the protection capability of facilities such as wire netting, baffle is not strong, and it is difficult to prevent animal invasion in special topography, and is difficult to in time discover after damaging, needs periodic maintenance. And the infrared mode is used for detection, so that the working capacity is lost possibly due to impact, topography and weather. The intrusion detection method provided by the scheme can completely solve the problem that equipment is damaged, the camera is deployed at a high position and is not easy to damage, and the equipment can be timely found through monitoring when damaged.
The portability is high: most of the existing target detection models cannot be well applied to various environmental hardware conditions, and due to the size and system version problems of the models, the model environments are possibly incompatible to be built and can only be applied to specific environments. The model provided by the invention can be suitable for various equipment conditions, and good detection precision and running delay can be obtained. For a software environment, the model of the invention can be operated in various Linux systems and Windows systems with stable versions; the model has the characteristics of light weight and low requirement on hardware, and can be deployed in various devices such as embedded development boards, mobile phones, computers and the like through transplanting.
Drawings
FIG. 1 is a flowchart of an anomaly target detection method based on a deep learning algorithm according to an embodiment of the present invention.
FIG. 2 is a schematic diagram of an anomaly target detection model based on a deep learning algorithm according to an embodiment of the present invention;
FIG. 3 is a graph showing comparison of pruning effects of a model according to an embodiment of the present invention;
FIG. 4 is a training flowchart of an anomaly object detection model based on a deep learning algorithm according to an embodiment of the present invention.
Detailed Description
The invention is further described below with reference to the drawings and examples. The following examples are only for more clearly illustrating the technical aspects of the present invention, and are not intended to limit the scope of the present invention.
In the description of the present invention, the meaning of a number is one or more, the meaning of a number is two or more, and greater than, less than, exceeding, etc. are understood to exclude the present number, and the meaning of a number is understood to include the present number. The description of the first and second is for the purpose of distinguishing between technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or implicitly indicating the precedence of the technical features indicated.
In the description of the present invention, the descriptions of the terms "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
Example 1
A power distribution station room anomaly detection method based on a deep learning algorithm comprises the following steps:
acquiring an image in a power distribution station room, and preprocessing the image, wherein the preprocessing comprises image scaling and image patching;
inputting the preprocessed image into a trained abnormal target detection model based on a deep learning algorithm;
determining an abnormal detection result of the power distribution station room according to the output of the abnormal target detection model;
the abnormal target detection model comprises an input layer, a main network with a CA attention mechanism, a neck network and an output layer; the backbone network is formed by combining a series of convolution layers, and a CA attention mechanism is used for fusing position information; the neck network adopts a structure of a feature map pyramid network FPN+ pixel aggregation network PAN, wherein the feature map pyramid network FPN is of a top-down network structure, high-level feature maps are transferred and fused through up-sampling operation, high-level strong semantic features are conveyed, convolution operation is carried out on the pixel aggregation network PAN from bottom to top, strong positioning features are transferred, and the detection capability of targets with different scales is improved; and combining the FPN with the PAN to obtain a final predicted feature map.
In some embodiments, the processing of the abnormal target detection model includes:
the image is sent into a backbone network through an input layer, and the feature extraction is carried out on the input image through the backbone network to generate a high-level feature map;
the neck network performs differentiation detection on the high-level feature images, performs cross-layer splicing through the up-sampling network and the convolutional neural network, and performs target detection on the feature images which are divided into different sizes;
and classifying the feature map output by the neck network through a convolution layer of the output layer to obtain an abnormal target detection result.
Further, in some embodiments, the division into different size feature maps performs different size object detection, including: a feature map for detecting the size of small targets 76 x 76, a feature map for detecting the size of medium targets 38 x 38 and a feature map for detecting the size of large targets 19 x 19 are obtained, respectively.
In some embodiments, the training method of the abnormal target detection model includes:
acquiring a training data set;
and inputting the training data set into an abnormal target detection model to be trained for training, carrying out weighted non-maximum suppression processing and confidence assessment on the output of the model, and carrying out optimization updating on model parameters by combining a loss function until a preset condition is reached, so as to obtain the trained abnormal target detection model.
Further, acquiring the training data set further comprises: and carrying out data enhancement processing on the training data set, wherein the data enhancement processing comprises Mosaic data enhancement, image overlapping and dropout noise addition processing.
Further, the mosaics are spliced by adopting 4 pictures in a random zooming, random cutting and random arrangement mode;
the image overlapping is to fuse two pictures with the same size according to random proportion to synthesize a new picture, and the proportion numerical values obey Beta distribution with parameters; the new picture is provided with the targets of the two pictures, and the labels of the two pictures are spliced together to form a new label;
the dropout noise is added to randomly generate a plurality of rectangular blocks with the same size on the image, so that the original picture is shielded, and the information of the positions of the rectangular blocks is lost.
Further, inputting the training data set into an abnormal target detection model to be trained for training, and carrying out weighted non-maximum suppression processing and confidence assessment on the output of the model, wherein the method comprises the following steps:
the output of the abnormal target detection model comprises the score, the confidence score and the position information of each category of the detected candidate target frame;
performing non-maximum value inhibition processing on each candidate target frame, and removing redundant target frames to obtain a high-quality target frame;
and comparing the confidence coefficient of the high-quality target frame with a preset confidence coefficient threshold value, judging whether the target frame is qualified or not in response to the fact that the confidence coefficient is larger than the preset confidence coefficient threshold value, storing the related target frame and the picture, and returning a result.
Further, the training method of the abnormal target detection model further comprises pruning the abnormal target detection model:
constraint is carried out by adding an L1 penalty term to a BN layer after a convolution layer in the neural network, so that model parameters are sparse; the output OUT formula of BN layer is as follows:
OUT＝γ*x i ′ +β
wherein, gamma is a scaling factor, and beta is a translation factor; x is x i ′ The input of the BN layer;
the gamma is normally distributed, a large number of gamma values tend to 0 by adding an L1 penalty term, and more layers which do not influence the result are removed;
wherein L is a loss function of model training, the first term in the formula (x, l (f (x, W, y) is the loss of model training, x is the input data, and the actual prediction result is calculated by the model weight W, the result f (x, W and the label result y calculate loss of loss, the second term lambda sigma) γ∈r gγ) is a constraint where γ is the polar radius, gγ) is the sum of the absolute values of the individual elements in the weight vector, λ is an adjustable regularization coefficient, and by setting the appropriate λ value, the weight is madeThe thinning effect is achieved again.
In some embodiments, the method comprises:
1. establishing a deep learning anomaly detection model
The design scheme of the invention emphasizes the efficiency and accuracy of detection. The former is realized by effectively reducing the parameters of the original algorithm and carrying out necessary simplification on the number of model channels on the premise of ensuring the correctness of the algorithm; the method introduces an attention mechanism algorithm, a data enhancement algorithm and other algorithms based on the original deep learning algorithm, and effectively solves the problem that abnormal behavior detection of the power distribution station room is positioned in classification based on deep learning.
Aiming at a power distribution station house scene, a lightweight abnormal target detection model is constructed, and the model can be divided into four parts of input, a backbone network, a neck network and output according to the current mainstream target detection model specification.
(1) An input end: at the input end of the model, the invention carries out data enhancement processing on the training data set.
Firstly, the data set is subjected to Mosaic data enhancement, and the Mosaic adopts 4 pictures to splice in a mode of random zooming, random cutting and random arrangement.
Second, an image overlay technique is added. The principle of image overlapping is to fuse two pictures with the same size according to random proportion to synthesize a new picture, and the proportion numerical values obey Beta distribution with parameters. The new picture has the targets of two pictures, and simultaneously the labels of the two pictures are spliced together to form a new label, so that the detection capability of the model can be effectively enhanced, the label containing noise data is resisted, and the robustness is improved.
Adding dropout noise is also a method of data enhancement. A plurality of rectangular blocks with the same size are randomly generated on the image, original pictures are shielded, information of positions of the original pictures is lost, model training complexity is increased, and 15% training samples are subjected to noise enhancement.
Because the detected target is in a very positive motion state, the shooting effect of the camera is poor, and the real-time image can be blurred. Blurring is performed on a plurality of pixels around the target so that the situation of animal shooting in a real scene is more similar.
(2) Backbone network: the main network is formed by combining a series of convolution layers, and the main function is to extract the characteristics of an input image and form high-dimensional characteristic diagrams with different granularities.
In the current convolutional neural network, it is difficult to obtain key features from global features. Simulating the ability of humans to quickly acquire key information from images, researchers have designed attention mechanisms. The invention embeds the CA attention mechanism into the backbone network of the YOLOv5s model to obtain better accuracy. The CA (Coordinate attention) attention mechanism can well integrate the position information, simultaneously consider the channel information and the position information of the feature space, and has better migration capability.
(3) Neck network: the neg part adopts a structure of FPN+PAN. The FPN is a top-down network structure, the high-level feature images are transferred and fused through the up-sampling operation, the high-level strong semantic features are conveyed, the PAN layer is subjected to convolution operation from bottom to top, the strong positioning features are conveyed, and the detection capability of targets with different scales is improved. And combining the two to obtain a final predicted feature map.
(4) And (3) outputting: and (3) performing a modeling box loss function calculation and a weighted nms non-maximum suppression process at the model output end, calculating a target most similar to a real target frame, outputting, and finally performing iterative updating on model parameters through back propagation and optimizer operation.
Through the model steps, an abnormal detection model with high precision and strong generalization capability can be obtained, but the detection speed and the model volume can be continuously optimized, so that the model efficiency is improved.
If the calculation amount can be effectively reduced by reducing the size of the input image, the accuracy of the model is affected, and the detection result of the model is seriously affected. Therefore, the network structure is cut by adopting a model pruning method, the model volume is reduced and the reasoning speed is increased while the accuracy is not reduced. First, constraint is performed by adding an L1 penalty term to BN (Batch Normalization) layers after a convolution layer in a neural network, so that model parameters are thinned. The output formula of BN layer is as follows:
OUT＝γ*x i ′ +β
where γ is the scaling factor and β is the translation factor. It can be seen that gamma is in direct proportion to output, if the value of gamma corresponding to a certain layer is too small, the output result can be ignored, and the influence of the layer on the detection result is proved to be small, and the removal of the layer does not influence the result precision. Under normal conditions, the magnitude of gamma is normally distributed, and a large amount of gamma values tend to 0 by adding an L1 penalty term, so that more layers which do not influence the result can be removed.
The first term in the formula is the loss of model training, the second term is the constraint condition, wherein ggamma) = |s|, lambda is an adjustable regularization coefficient, and the weight achieves the thinning effect by setting a proper lambda value. After sparsification, relevant layers with small sparsity parameters are cut off, the influence of the layers on model precision is small, and the performance of the model can be ensured after cutting. After the result pruning iterative training in the text, the detection time delay and the memory occupation of the model are reduced.
So far, after model training is finished and construction is finished, the model can be directly subjected to anomaly detection after deployment.
2. Deployment detection using training models
The trained model is deployed on a computer, a server or terminal equipment and connected with a camera, so that real-time abnormal target detection can be performed.
The method comprises the following steps:
the cameras distributed in each station room monitor the real-time conditions in the station room in real time, and when a detection switch is turned on, the model is activated and the real-time detection is carried out:
firstly, an image shot by a camera in a station house is sent into a model for preprocessing, and parameters of the image are changed to facilitate model detection, and the method comprises the following steps: image scaling, image inpainting, etc.; the invention adds CA attention mechanism in the main Network, which can better focus on the target difficult to identify caused by light, angle and other factors, the generated feature map is differentiated and detected by the Neck Network (Neck Network), the feature map is divided into feature maps with different sizes by up-sampling, residual structure and concat splicing form to detect targets with different sizes, so as to improve the detection precision. Finally, the images (including cats, dogs, falling personnel and personnel not matched with the safety helmet) are respectively classified through a convolution layer (Convolutional layer), and the identification result is returned. Because the returned result is in a tensor form, certain post-processing is needed to convert the returned result into the identification image which is wanted by the user, and meanwhile, the confidence and classification result can be printed. Finally, comparing the value of the initially set confidence threshold value with the value of the finally detected threshold value: if the detected abnormal confidence is smaller than the threshold value, the result is not qualified and marked and stored; if the confidence coefficient is larger than the threshold value, judging that the target frame and the picture are qualified, storing the related target frame and the picture, and returning a result.
The specific steps of the abnormal target detection are as follows:
step 1: initializing-a set confidence threshold conf-thres, model file weights, detecting parameters such as an image path source, an image preprocessing size img-size and the like;
step 2: the image obtains 1024 feature images with the same size through a model Backbone Network (backhaul Network) comprising two layers of CA attention mechanisms;
step 3: the feature map is characterized in that a large number of up-sampling networks and convolutional neural networks are adopted in a Neck Network (Neck Network), and cross-layer splicing is carried out, so that feature maps for detecting the sizes of small targets 76, medium targets 38 and 19 x 19 of large targets are obtained respectively;
step 4: the feature map is sent to a final layer of convolution layer for classification and output, and the score, the confidence score and the position information of each category are respectively obtained for each detected target frame;
step 5: performing NMS (Non Maximum Suppression) non-maximum value inhibition processing on each target frame to remove redundant target frames;
step 6: and outputting and displaying the rest high-quality target frames on the graph, and simultaneously marking out confidence scores and category information and storing the results.
Example 2
In a second aspect, the present embodiment provides a substation room anomaly detection device based on a deep learning algorithm, including a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is operative according to the instructions to perform the steps of the method according to embodiment 1.
Example 3
In a third aspect, the present embodiment provides a storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method described in embodiment 1.
Example 4
In a fourth aspect, the present invention provides a computing device comprising:
one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods described in embodiment 1.
It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The foregoing is only a preferred embodiment of the invention, it being noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the invention.
Claims (10)
1. The power distribution station room anomaly detection method based on the deep learning algorithm is characterized by comprising the following steps of:
acquiring an image in a power distribution station room, and preprocessing the image, wherein the preprocessing comprises image scaling and image patching;
inputting the preprocessed image into a trained abnormal target detection model based on a deep learning algorithm;
determining an abnormal detection result of the power distribution station room according to the output of the abnormal target detection model;
the abnormal target detection model comprises an input layer, a main network with a CA attention mechanism, a neck network and an output layer; the backbone network is formed by combining a series of convolution layers, and a CA attention mechanism is used for fusing position information; the neck network adopts a structure of a feature map pyramid network FPN+ pixel aggregation network PAN, wherein the feature map pyramid network FPN is of a top-down network structure, high-level feature maps are transferred and fused through up-sampling operation, high-level strong semantic features are conveyed, convolution operation is carried out on the pixel aggregation network PAN from bottom to top, strong positioning features are transferred, and the detection capability of targets with different scales is improved; and combining the FPN with the PAN to obtain a final predicted feature map.
2. The method for detecting abnormal conditions of a substation room based on the deep learning algorithm according to claim 1, wherein the processing procedure of the abnormal target detection model comprises:
the image is sent into a backbone network through an input layer, and the feature extraction is carried out on the input image through the backbone network to generate a high-level feature map;
the neck network performs differentiation detection on the high-level feature images, performs cross-layer splicing through the up-sampling network and the convolutional neural network, and performs target detection on the feature images which are divided into different sizes;
and classifying the feature map output by the neck network through a convolution layer of the output layer to obtain an abnormal target detection result.
3. The method for detecting abnormal conditions of a substation room based on a deep learning algorithm according to claim 2, wherein the method for detecting targets of different sizes by dividing the characteristic map of different sizes comprises the steps of: a feature map for detecting the size of small targets 76 x 76, a feature map for detecting the size of medium targets 38 x 38 and a feature map for detecting the size of large targets 19 x 19 are obtained, respectively.
4. The method for detecting abnormal conditions of a substation room based on a deep learning algorithm according to claim 1, wherein the training method of the abnormal target detection model comprises the following steps:
acquiring a training data set;
and inputting the training data set into an abnormal target detection model to be trained for training, carrying out weighted non-maximum suppression processing and confidence assessment on the output of the model, and carrying out optimization updating on model parameters by combining a loss function until a preset condition is reached, so as to obtain the trained abnormal target detection model.
5. The method for detecting abnormal conditions in a substation room based on a deep learning algorithm according to claim 4, wherein acquiring the training data set further comprises: and carrying out data enhancement processing on the training data set, wherein the data enhancement processing comprises Mosaic data enhancement, image overlapping and dropout noise addition processing.
6. The power distribution station house anomaly detection method based on the deep learning algorithm according to claim 5, wherein the mosaics data enhancement adopts 4 pictures to splice in a mode of random scaling, random cutting and random arrangement;
the image overlapping is to fuse two pictures with the same size according to random proportion to synthesize a new picture, and the proportion numerical values obey Beta distribution with parameters; the new picture is provided with the targets of the two pictures, and the labels of the two pictures are spliced together to form a new label;
the dropout noise is added to randomly generate a plurality of rectangular blocks with the same size on the image, so that the original picture is shielded, and the information of the positions of the rectangular blocks is lost.
7. The method for detecting abnormal power distribution station house based on deep learning algorithm according to claim 4, wherein the step of inputting the training data set into an abnormal target detection model to be trained for training, and performing weighted non-maximum suppression processing and confidence evaluation on the output of the model comprises the steps of:
the output of the abnormal target detection model comprises the score, the confidence score and the position information of each category of the detected candidate target frame;
performing non-maximum value inhibition processing on each candidate target frame, and removing redundant target frames to obtain a high-quality target frame;
and comparing the confidence coefficient of the high-quality target frame with a preset confidence coefficient threshold value, judging whether the target frame is qualified or not in response to the fact that the confidence coefficient is larger than the preset confidence coefficient threshold value, storing the related target frame and the picture, and returning a result.
8. The method for detecting abnormal conditions in a substation room based on a deep learning algorithm according to claim 4, wherein the training method for the abnormal target detection model further comprises pruning the abnormal target detection model:
constraint is carried out by adding an L1 penalty term to a BN layer after a convolution layer in the neural network, so that model parameters are sparse; the output OUT formula of BN layer is as follows:
OUT＝γ*x i ′ +β
wherein, gamma is a scaling factor, and beta is a translation factor; x is x i ′ The input of the BN layer;
the gamma is normally distributed, a large number of gamma values tend to 0 by adding an L1 penalty term, and more layers which do not influence the result are removed;
wherein L is a loss function of model training, the first term in the formula (x, l (f (x, W, y) is the loss of model training, x is the input data, and the model weight W calculates the realThe inter-prediction result, the result f (x, W and the label result y calculate loss of loss, the second term lambda sigma γ∈r gγ) is a constraint condition, where γ is a polar radius, gγ) is a sum of absolute values of elements in a weight vector, λ is an adjustable regularization coefficient, and by setting a proper λ value, the weight achieves a sparsifying effect.
9. The power distribution station room abnormality detection device based on the deep learning algorithm is characterized by comprising a processor and a storage medium;
the storage medium is used for storing instructions;
the processor being operative according to the instructions to perform the steps of the method according to any one of claims 1 to 8.
10. A storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the method according to any of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310049195.8A CN116052082A (en) | 2023-02-01 | 2023-02-01 | Power distribution station room anomaly detection method and device based on deep learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310049195.8A CN116052082A (en) | 2023-02-01 | 2023-02-01 | Power distribution station room anomaly detection method and device based on deep learning algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116052082A true CN116052082A (en) | 2023-05-02 |
Family
ID=86116459
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310049195.8A Pending CN116052082A (en) | 2023-02-01 | 2023-02-01 | Power distribution station room anomaly detection method and device based on deep learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116052082A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116681980A (en) * | 2023-07-31 | 2023-09-01 | 北京建筑大学 | Deep learning-based large-deletion-rate image restoration method, device and storage medium |
| CN117468084A (en) * | 2023-12-27 | 2024-01-30 | 浙江晶盛机电股份有限公司 | Crystal bar growth control method and device, crystal growth furnace system and computer equipment |
-
2023
- 2023-02-01 CN CN202310049195.8A patent/CN116052082A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116681980A (en) * | 2023-07-31 | 2023-09-01 | 北京建筑大学 | Deep learning-based large-deletion-rate image restoration method, device and storage medium |
| CN116681980B (en) * | 2023-07-31 | 2023-10-20 | 北京建筑大学 | Deep learning-based large-deletion-rate image restoration method, device and storage medium |
| CN117468084A (en) * | 2023-12-27 | 2024-01-30 | 浙江晶盛机电股份有限公司 | Crystal bar growth control method and device, crystal growth furnace system and computer equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111353413B (en) | Low-missing-report-rate defect identification method for power transmission equipment | |
| Jiao et al. | A yolov3-based learning strategy for real-time uav-based forest fire detection | |
| CN110084165B (en) | Intelligent identification and early warning method for abnormal events in open scene of power field based on edge calculation | |
| CN110688925B (en) | Cascade target identification method and system based on deep learning | |
| CN116052082A (en) | Power distribution station room anomaly detection method and device based on deep learning algorithm | |
| KR20180135898A (en) | Systems and methods for training object classifiers by machine learning | |
| Maksymiv et al. | Real-time fire detection method combining AdaBoost, LBP and convolutional neural network in video sequence | |
| CN111553403B (en) | Smog detection method and system based on pseudo-3D convolutional neural network | |
| CN115457428A (en) | Improved YOLOv5 fire detection method and device integrating adjustable coordinate residual attention | |
| CN113642474A (en) | Hazardous area personnel monitoring method based on YOLOV5 | |
| Jiang et al. | A self-attention network for smoke detection | |
| CN111539325A (en) | Forest fire detection method based on deep learning | |
| CN115546742A (en) | Rail foreign matter identification method and system based on monocular thermal infrared camera | |
| Yandouzi et al. | Investigation of combining deep learning object recognition with drones for forest fire detection and monitoring | |
| CN114155551A (en) | Improved pedestrian detection method and device based on YOLOv3 under complex environment | |
| CN112270671B (en) | Image detection method, device, electronic equipment and storage medium | |
| CN113191273A (en) | Oil field well site video target detection and identification method and system based on neural network | |
| CN117423157A (en) | Mine abnormal video action understanding method combining migration learning and regional invasion | |
| CN112329550A (en) | Weak supervision learning-based disaster-stricken building rapid positioning evaluation method and device | |
| CN116884192A (en) | Power production operation risk early warning method, system and equipment | |
| CN116740495A (en) | Training method and defect detection method for defect detection model of road and bridge tunnel | |
| CN112784788B (en) | Low-resolution target detection method | |
| Jandhyala et al. | Forest Fire Classification and Detection in Aerial Images using Inception-V3 and SSD Models | |
| Madhumathi et al. | Advanced Wild Animal Detection and Alert System using YOLO V5 Model | |
| Lestari et al. | Comparison of two deep learning methods for detecting fire |
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 | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Zhang Hao Inventor after: Zhu Tianze Inventor after: Jiang Chao Inventor after: Zhao Cheng Inventor after: Chen Zhiming Inventor after: Li Chunpeng Inventor after: Wang Baoping Inventor after: Luan Qiqi Inventor after: Yang Xiaoping Inventor after: Li Jun Inventor after: Guan Guofei Inventor after: Song Qingwu Inventor after: Jiang Feng Inventor before: Zhang Hao Inventor before: Zhu Tianze Inventor before: Jiang Chao Inventor before: Zhao Cheng Inventor before: Chen Zhiming Inventor before: Su Yubiao Inventor before: Jiang Lincen Inventor before: Xu He Inventor before: Ji Yimu Inventor before: Liu Shangdong Inventor before: Li Chunpeng Inventor before: Wang Baoping Inventor before: Luan Qiqi Inventor before: Yang Xiaoping Inventor before: Li Jun Inventor before: Guan Guofei Inventor before: Song Qingwu Inventor before: Jiang Feng |
|
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20240122 Address after: 211100 No. 19 Tianyuan Middle Road, Jiangning Science Park, Nanjing City, Jiangsu Province Applicant after: JIANGSU FANGTIAN POWER TECHNOLOGY Co.,Ltd. Country or region after: China Address before: 211100 No. 19 Tianyuan Middle Road, Jiangning Science Park, Nanjing City, Jiangsu Province Applicant before: JIANGSU FANGTIAN POWER TECHNOLOGY Co.,Ltd. Country or region before: China Applicant before: NANJING University OF POSTS AND TELECOMMUNICATIONS |