CN110334661A - Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning - Google Patents
Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning Download PDFInfo
- Publication number
- CN110334661A CN110334661A CN201910612845.9A CN201910612845A CN110334661A CN 110334661 A CN110334661 A CN 110334661A CN 201910612845 A CN201910612845 A CN 201910612845A CN 110334661 A CN110334661 A CN 110334661A
- Authority
- CN
- China
- Prior art keywords
- training
- network
- infrared
- deep learning
- algorithm
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/088—Aspects of digital computing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
-
- 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
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to the often fever point target detecting method of the infrared defeated variation based on deep learning is planted, includes the following steps: step 1, to site infrare abnormal heating image data sample collection, establish training dataset；Step 2 is constructed prototype network, is trained using training dataset to the prototype network established, obtains network model；Step 3, the network model established using training, infrared image to be identified is identified, obtain the result of infrared picture abnormal heating point identification and positioning, it is verified, the abnormal heating fault point identification in infrared image can be oriented with higher accuracy rate from more complicated detection background, have preferable detection effect, method provided by the present invention can detect for power transmission and transforming equipment infrared detection fault point intelligent positioning and provide more accurate efficient detection thinking.
Description
Technical field
The present invention relates to a kind of infrared power transmission and transformation abnormal heating point target detecting method based on deep learning.
Background technique
Guarantee the safe and stable operation of power system device, reduce the expense of operation with maintenance, promotes fortune inspection efficiency and standard
True rate is current new challenge.The operational mode of domestic power grid is broadly divided into three kinds: unattended mode, someone's attended mode with
And few people on duty mode.Wherein someone's attended mode is traditional substation operation mode, and unattended mode is a kind of new
Operational mode.There are the following problems for traditional repair and maintenance: (1) needing to carry out periodic interruption maintenance, with power transformation
Stand and large capacity, super-pressure power equipment investment, currently running and service personnel do not adapt to power network development
Needs；(2) scheduled outage service work, which relies primarily on, periodically carries out insulation preventive trial to equipment, and needs to have a power failure
Maintenance, cannot reflect insulation status of the power system device in normal charging operation comprehensively；(3) scheduled outage is overhauled
Heavy workload exerts a certain influence for power grid and user, and be easy to cause safety accident.Therefore operation of power networks maintenance
Change it is extremely urgent, especially under the influence of big data artificial intelligence power grid, how to improve the intelligent operation of power grid
How required level of service makes full use of the human resources of power grid, and shortening fault time makes power grid high-efficiency operation particularly significant.
The infrared anomaly febrile state monitoring and fault diagnosis of power transmission and transforming equipment stablizes to Guan Chong power system security
It wants.For a long time, many domestic and foreign scholars have been devoted to research power equipment abnormal heating method for diagnosing faults, propose as
The detection methods such as voltage's distribiuting method, impulse electric corona current method, electric field measurement method, infrared imaging method.Wherein, infrared imaging method is logical
The temperature height for crossing the radiation effect reaction object under test of infrared band, has non-contact, not damaged, high reliablity, thermometric face
Accumulate the advantages such as big.
Nevertheless, infrared monitoring equipment is multi-point and wide-ranging due to a variety of detection platforms, the infrared picture data product of magnanimity
Pressure, it is difficult to which timely and effective treatment effeciency is low etc., these information excavatings and target identification all to inspection image propose higher
It is required that.Traditional intelligent positioning technology is accorded with using the feature of engineer, such as Adaboost, edge detection, in conjunction with texture, face
The shallow-layers feature such as color, shape is identified.These algorithms have the defects of accuracy rate is low, generalization ability is poor, and model ossifys.With
The development of machine vision, some have the intelligent Model of Target Recognition for adjusting ginseng and sample adaptive learning and come into being.
Artificial intelligence (Artificial Intelligent) is a research, develops for simulating, extending and extending people
Intelligence theory, method, a new technological sciences of technology and application system.Machine learning (Machine
Learning), a method of realizing artificial intelligence.Core of the machine learning as artificial intelligence is a multi-field intersection
Subject is related to the multiple subjects such as probability theory, statistics.Deep learning (Deep Learning), as the core for realizing machine learning
Heart key technology.It is transported as theory of algorithm emerging in recent years in power grids such as image classification, target identification, Target Segmentations
Several scenes application is examined, a collection of breakthrough technological achievement has gradually been emerged.Triadic relation deepens layer by layer.
The realization of artificial intelligence deep learning needs big data as support in power industry, and the correlation of deep learning is ground
Study carefully and applies also first meeting clue.In operation of power networks status monitoring, electric power analysis load prediction, user behavior analysis prediction, electric power
Gradually start to play effect and effect in terms of system trouble analysis, Fault Diagnosis for Electrical Equipment.
For power transmission and transforming equipment intellectualized detection equipment carrying platform with Intelligent Mobile Robot platform and power transmission line
Based on the unmanned aerial vehicle platform of road.Wherein, intelligent inspection robot is using mobile robot as carrier, with visible light camera, red
Other detecting instruments such as outer thermal imaging system can be realized short distance observation device as load system, need to pass through deep learning
Algorithm realizes the positioning identified to barrier, to the functions such as transmission line of electricity and its autonomous inspection of line corridor, to improve
The accuracy rate of power grid fortune inspection.UAV flight's video camera, has the function of high accuracy positioning and automatic camera, can be in time by power grid
Running relevant device state picture is transmitted to computer, by deep learning algorithm to collected visible light, it is infrared heat at
Picture, UV corona image video analysis, and then judge equipment state.
Basic model inside artificial intelligence deep learning has been roughly divided into 3 classes: multiple perceptron model；Depth nerve net
Network model and recurrent neural networks model.Its representative is DBN (Deep belief network) deepness belief network, CNN respectively
(Convolution Neural Networks) convolutional neural networks, RNN (Recurrent neural network) recurrence mind
Through network.
2006, Geoffrey Hinton proposed deepness belief network (DBN) and its efficient learning algorithm, that is, passed through
Pre-training and the method (Pre-training+Fine-tuning) of later period training fine tuning carry out, from largely avoiding ladder
The case where degree disappears, which delivers to " Science " periodical, becomes one of deep learning algorithm main method thereafter.
DBN pattern-recognition scene with high-dimensional feature amount suitable for the fortune inspection of electric power power transmission and transforming equipment, itself can
There is more variant according to different concrete application scenes, its improvement focuses primarily upon its improvement for forming " part " RBM, has
Convolution DBN (CDBN) and other algorithm connected applications etc..Such as in the literature, after establishing characteristic of transformer gas sample collection, benefit
With depth confidence network (DBN) method in deep neural network, depth of assortment confidence network diagnosis model is constructed, it is final real
The on-line fault diagnosis of transformer is showed.Document uses the circuit breaker failure diagnosis side that DBN and Softmax classifier combines
Method, extracting high-level characteristic information using depth confidence network reduces data dimension, eventually by Softmax to fault category
Carry out classification diagnosis assessment.Document proposes a kind of partial discharge of transformer mode identification method based on DBN, and using adaptive
Learning rate is answered to control it during model training to the optimizing ability of globally optimal solution, realizes higher recognition accuracy
With the faster algorithm speed of service.
Convolutional neural networks are artificial neural network processing most common one kind of two-dimensional image information, it has also become present image
The research hotspot and mainstream of classification and identification field.The advantage shows the most obvious when the input of network is multidimensional image,
Make image directly as the input quantity of network model can avoid that human subjective is needed to set complicated feature in tional identification algorithm
Extract the process with data reconstruction.
CNN inputs the applied field of this kind of two-dimensions input parameter suitable for the fortune inspection of electric power power transmission and transforming equipment with image/video
Scape.CNN algorithm is itself initially applied to the classification and prediction scene of image or video frame, in recent years researcher couple
CNN has carried out various mutation and optimization, makes it possible in the more extensive electric power equipment such as target identification, Target Segmentation
It is applied in the scene of vision.As author uses the target identification branch Faster R-CNN model of CNN to power transmission line in document
The failure of road insulator carries out fault identification and detection.In document, author is by input cable warping apparatus picture to CNN's
YOLO model is improved, realizes the positioning and identification of cable machinery abnormality location point.
Since DNN and CNN neural network is there is that can not model time series, still, time series
It is undivided, has between input time axis and input information multiple for the voice and video information being closely related with time shaft
Miscellaneous relevance.In order to cope with such demand, be born RNN Recognition with Recurrent Neural Network.As shown in figure 3, RNN Recognition with Recurrent Neural Network
The characteristics of to be closely related with time series, input layer upper element in sequence hides working together to for layer signal
Current hidden layer, and it is sequentially delivered to the last layer, solve the classification and forecasting problem of time series data.
Researcher has also carried out correlative study and application in terms of electric power inspection for RNN recurrent neural network.Such as
Document uses the diagnosing fault of power transformer model that Recognition with Recurrent Neural Network and bat algorithm combine, and passes through bat algorithm
The parameter for optimizing Recognition with Recurrent Neural Network, realizes the diagnosis and assessment of transformer fault.Document is used based on after timing decomposition
It is realized to the Recognition with Recurrent Neural Network prediction model of propagation algorithm in conjunction with a variety of external factor characteristic parameters to the pre- of power load
It surveys, correlated results shows that prediction added the conventional method that compares with a degree of promotion.
With the development of artificial intelligence deep learning theoretical system, Ross Girshick team is proposed within 2015
Faster-RCNN model is won the championship at one stroke in COCO detection contest, realizes the double excellent prominent of target detection accuracy rate and test speed
It is broken.The algorithm is improved on the basis of RCNN and Fast-RCNN, is introduced region and is suggested network (Region
Proposal Network, RPN) it is used to extract detection zone, it will test within step unification to a depth network frame, greatly
Detection speed and efficiency are improved greatly.Currently, Faster-RCNN model is applied to grinding for power equipment abnormal heating targeted diagnostics
Study carefully also in the elementary step.
Summary of the invention
The purpose of the present invention is to overcome the deficiencies of existing technologies, and provides a kind of infrared defeated variation based on deep learning
Often fever point target detecting method.
To achieve the goals above, the technical solution used in the present invention is:
A kind of infrared defeated normal fever point target detecting method of variation based on deep learning, includes the following steps:
Step 1 establishes training dataset to site infrare abnormal heating image data sample collection；
Step 2 is constructed prototype network, is trained using training dataset to the prototype network established, obtains network
Model；
Step 3, the network model established using training, identifies infrared image to be identified, obtains infrared picture
The result of abnormal heating point identification and positioning.
As some embodiments of the present invention, the sample collection method of the step 1 is as follows: being marked by data soft
Frame choosing is marked to heat generating spot in the infrared failure picture with abnormal heating in part, arranges infrared as correspondence for xml document
The label of picture training.
As some embodiments of the present invention, the prototype network of the step 2 uses four step change training methods, specifically
Include:
1) network RPN pre-training is suggested in region: using the training dataset of step 1 RPN network being carried out by ZF network
There is supervision pre-training；
2) convolutional neural networks of ZF model Fast R-CNN network pre-training: are carried out using the training dataset of step 1
Have supervision pre-training, using its training at the end of the network parameter that generates as the initiation parameter of joint training；
3) RPN network fine tuning training: the positive sample of RPN network fine tuning training derives from and handmarking's frame (Ground
Truth the data sample that alternative frame (anchors) registration (IoU) is greater than 70%) is generated with algorithm, negative sample, which derives from, to be less than
The algorithm of 30% registration generates alternative frame, and positive sample only indicates prospect, and negative sample only indicates background；Operation is returned only for just
Sample carries out；All algorithms beyond image boundary are abandoned when training generates alternative frame, it is raw beyond the algorithm behind boundary to removing
At alternative frame collection use non-maxima suppression；Fixed shared convolutional layer, that is, it is zero that its learning rate, which is arranged, more without parameter
Newly, the network parameter of the exclusive layer of RPN network is only finely tuned, realizes shared convolution；
4) Fast R-CNN network fine tuning training: the positive sample of Fast R-CNN network fine tuning training derives from handmarking
Frame and algorithm generate the data sample that alternative frame registration is greater than threshold region suggestion, and negative sample is from handmarking's frame and calculates
Method generates the data sample that alternative frame registration is less than threshold region suggestion, only finely tunes the exclusive layer of Fast R-CNN, so far shape
At the Unified Network of shared convolution model.
In certain embodiments of the present invention, the threshold value is set as 0.4.
In certain embodiments of the present invention, network RPN pre-training is suggested in the region, especially by backpropagation
(back-propagation, BP) and stochastic gradient descent (stochastic gradient descent, SGD) carry out end and arrive
End training trains this network according to " image-centric " sampling policy in FastR-CNN, and each minimum step is by wrapping
The single image composition of many positive negative samples is contained, while using 0.01 mean value of standard deviation is 0 Gaussian Profile to newly-increased layer
Random initializtion.
In certain embodiments of the present invention, training rate is 0.0001.
The beneficial effects of adopting the technical scheme are that
Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning proposed by the invention, it is verified,
The abnormal heating fault point in infrared image can be identified by positioning with higher accuracy rate from more complicated detection background
Out, has preferable detection effect.
Method provided by the present invention can be detected for power transmission and transforming equipment infrared detection fault point intelligent positioning and be provided more
Precisely efficient detection thinking.
Detailed description of the invention
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art are briefly described.
Fig. 1 is model combined training flow chart；
Fig. 2 is confidence threshold value and accuracy rate relational graph；
Fig. 3 is model alternately training first step training curve；
Fig. 4 is model alternately training second step training curve；
Fig. 5 is model alternately training third step training curve；
Fig. 6 model alternately trains the 4th step training curve；
(A is infrared original image to Fig. 7 recognition effect comparison diagram, B is the method for the present invention recognition effect figure, C is conventional pixel threshold value
Recognition effect figure)；
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, invention is carried out combined with specific embodiments below
Clear, complete description.
Embodiment 1
Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning is managed using artificial intelligence deep learning
By Faster-RCNN algorithm model improved in system, by carrying out to site infrare abnormal heating image data sample collection
The training and verifying of algorithm model, it is final to realize power transmission and transforming equipment infrared image abnormal heating failure point target detection.
Step 1: the arrangement of data and label
The training dataset of power transmission and transforming equipment infrared imaging failure picture library composition, which has the following characteristics that, specifically includes iron
Red, rainbow, gray scale, ashen etc.；Trade mark, aiming frame, temperature value of infrared thermoviewer in infrared picture etc. is a variety of simultaneously
Information can be superimposed upon on infrared picture；Infrared picture is to be taken on site to collect, and background includes various other power equipments, interference
It is more serious；Site infrare thermal imaging system shooting angle is not fixed, there are a variety of shooting angle, the defect power equipment that is taken
Shape feature difference is larger, and the power equipment of abnormal heating failure includes suspension insulator, support insulator, disconnecting link, conducting wire, gold
Tool etc., building is 1270 total, is carried out by data marking software to heat generating spot in the infrared failure picture with abnormal heating
The choosing of handmarking's frame arranges the label for xml document as corresponding infrared picture training.
Step 2: prototype network combined training
Faster R-CNN algorithm is in abnormal heating point infrared image detection algorithm principle frame in deep learning algorithm system
Frame figure, first to the ZF depth convolution net being made of 13 layers of volume collection layer, 13 layers Relu layers (amendment linear unit), 4 layers of pond layer
On the one hand network mode input arbitrary size picture is input to by ZF network propagated forward to finally shared convolution characteristic pattern
Candidate region network obtains position coordinates and classification score information, on the other hand merges in the pond ROI layer with convolution characteristic pattern,
The feature that corresponding region is suggested is extracted, the full articulamentum of multilayer is eventually passed through and exports the classification score in the region and detect detection and determine
The classification score of output and coordinate frame position are superimposed upon on input picture by the coordinate frame of position, and output, which is shown, inputs infrared picture
The result of fault point identification and positioning.
The basis of the successful training of model is the gradually convergence of loss function during model training, FasterRCNN's
Target loss function such as formula (1) is as follows:
Wherein target loss function include Classification Loss and return to loss two parts.In loss function, i indicates one
The subscript of some reference windows (Anchor), p in minimum training pace (mini-batch)iIndicate that i-th of reference windows are
The probability for detecting target, when reference windows are negative sampleWhen reference windows are positive sampleIt follows that
Return loss item only can just be activated in the case that when reference windows are positive sample.ti={ tx,ty,tw,thIt is a vector,
Indicate the x coordinate, y-coordinate, width, highly this quadrinomial parameter coordinate of the encirclement frame (boundingbox) of the prediction,Be with
The coordinate vector of the callout box (groundtruthboundingbox) of positive sample reference windows.It is one two points
Shown in class loss function such as formula (2):
It is to return shown in loss function such as formula (3):
Wherein shown in the definition of R function such as formula (4):
Lambda parameter is used to weigh Classification Loss LclsL is lost with returning toreg, default value λ=10；NclsAnd NregIt is respectively intended to mark
Standardization Classification Loss item LclsItem L is lost with returningreg, default and set with minimum training pace value (mini-batch size) for 256
Set Ncls, N is initialized with the range intervals of reference windows position numerical value to 2400reg。
Since Faster R-CNN includes that two network structure parts of network are suggested in Fast R-CNN and region.RPN and
Fast R-CNN is stand-alone training, differently to modify their convolutional layer.This is not only to define one to include
The individual networks of RPN and Fast R-CNN, then with backpropagation combined optimization, it is so simple.Therefore it needs to develop one kind
Allow to share the technology of convolutional layer between two networks, rather than learns two networks respectively.The reason is that Fast R-CNN training according to
Lai Yu fixed target Suggestion box, when changing simultaneously proposed mechanism, learning model Fast R-CNN may not restrain.Such as figure
Shown in 1, the present invention replaces training method by 4 steps, continues to optimize study to sharing feature, each step is for frequency of training
1000 times, learning rate 0.001.
First step training: suggest network RPN pre-training in region: using the infrared picture of abnormal heating described in 1.2 trifles
Sample set makes RPN network carry out supervision pre-training by ZF network, especially by backpropagation (back-propagation,
BP) and stochastic gradient descent (stochastic gradient descent, SGD) carries out end-to-end training.According to FastR-
" image-centric " sampling policy in CNN trains this network.Each minimum step is by containing many positive negative samples
Single image composition.Using 0.01 mean value of standard deviation simultaneously is 0 Gaussian Profile to newly-increased layer random initializtion.
Second step training: the convolution of ZF model is carried out using the infrared picture sample collection of abnormal heating described in 1.2 trifles
Neural network has supervision pre-training, and the network parameter generated at the end of being trained using it is joined as the initialization of joint training
Number.
Third step training: the fine tuning training of RPN network: the positive sample of RPN network fine tuning training derives from and handmarking's frame
(Ground Truth) and algorithm generate the data sample that the alternative frame of anchor (anchors) registration (IoU) is greater than 70%, negative sample
Alternative frame is generated from the algorithm less than 30% registration.Positive sample only indicates prospect, and negative sample only indicates background；Return behaviour
Make to carry out only for positive sample；All algorithms beyond image boundary are abandoned when training and generate alternative frame, otherwise in training process
In can generate larger reluctant amendment error term, cause training process that can not restrain；To the algorithm removed after exceeding boundary
The alternative frame collection generated uses non-maxima suppression.Fixed shared convolutional layer, that is, it is zero that its learning rate, which is arranged, without parameter
It updates, only finely tunes the network parameter of the exclusive layer of RPN network, realize shared convolution.
4th step training: Fast R-CNN network fine tuning training: the positive sample source of Fast R-CNN network fine tuning training
The data sample that alternative frame registration is greater than threshold region suggestion is generated in handmarking's frame and algorithm, negative sample is from artificial
Indicia framing and algorithm generate the data sample that alternative frame registration is less than threshold region suggestion, and it is exclusive only to finely tune Fast R-CNN
Layer so far forms the Unified Network of shared convolution model.
Step 3: identification
The network model established using training, identifies infrared image to be identified, obtains infrared picture and sends out extremely
The result of hot spot identification and positioning.
The analysis of 1 detection error of test example
Test discovery is carried out to trained model, different threshold values is affected to testing result, first right here
Detection confidence threshold value is discussed.When detecting the confidence value in posting greater than threshold value, test picture tool is represented
There is abnormal heating point, be marked using posting, and the mesh that the heat generating spot in posting arrives for prototype network institute fixation and recognition
Mark；When detecting the confidence value in posting and being less than threshold value, then representing the test picture does not have an abnormal heating point, i.e., this is fixed
Position frame, which is ignored, to be disregarded.
24 infrared anomaly heating accident test images are randomly selected to carry out not having to the test under threshold value, infrared anomaly fever
The accuracy rate of fault identification is collectively formed by the false detection rate and omission factor identified, is arranged under different confidence threshold values, abnormal heating
The accuracy rate of point identification is as shown in Figure 3.The result shows that false detection rate and omission factor are opposite when confidence threshold value is set as 0.4
Equilibrium, is held at relatively low level, and the effect of recognition detection is best.
The analysis of 2 deep learning algorithm model training parameter of test example
In deep learning algorithm model training process, loss rate score (Loss) play the role of it is very important,
Whether the size of numerical value restrains and convergent effect as the fluctuation of frequency of training reflects model.At four of model training
In key step, training rate has selected tri- kinds of numerical value of typical 0.001/0.0001/0.00001, the training speed of each step
The training curve of rate and loss rate score is shown in attached drawing 3-6.When selecting higher trained rate value i.e. 0.001, such as attached drawing 4,5 and
Shown in 6, for training curve since training rate is very fast, training pace is larger, and fluctuation and unstability are also more obvious, this is all
Model inspection effect can finally be influenced；When selecting lower trained rate value i.e. 0.00001, as shown in attached drawing 5 and 7, by
Slower in training rate, training, training pace is smaller, is easily trapped into regional area optimal value rather than global optimum, cannot
It realizes lower loss late, finally prevents model from effectively carrying out recognition detection to target.For synthesis, training rate is selected
It is optimal selection when being 0.0001.
Effect example
Using method provided by the present invention and traditional algorithm for detecting heat generating spot based on pixel threshold respectively to having
The infrared image of interference is identified, is difficult to effective identify by the visible conventional method of attached drawing 7 and is partitioned into practical heat generating spot, and depth
Learning algorithm model can be with the position of heat generating spot at significant notation.
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify to technical solution documented by previous embodiment or equivalent replacement of some of the technical features；And
These are modified or replaceed, the spirit and model of technical solution of the embodiment of the present invention that it does not separate the essence of the corresponding technical solution
It encloses.
Claims (6)
- The point target detecting method 1. a kind of infrared defeated variation based on deep learning is often generated heat, which is characterized in that including walking as follows It is rapid:Step 1 establishes training dataset to site infrare abnormal heating image data sample collection；Step 2 is constructed prototype network, is trained using training dataset to the prototype network established, obtains network mould Type；Step 3, the network model established using training, identifies infrared image to be identified, and it is abnormal to obtain infrared picture The result of heat generating spot identification and positioning.
- The point target detecting method 2. a kind of infrared defeated variation based on deep learning according to claim 1 is often generated heat, It is characterized in that, the sample collection method of the step 1 is as follows: by data marking software to the infrared event with abnormal heating Frame choosing is marked in heat generating spot in barrier picture, arranges the label for xml document as corresponding infrared picture training.
- The point target detecting method 3. a kind of infrared defeated variation based on deep learning according to claim 2 is often generated heat, It being characterized in that, the prototype network of the step 2 uses four step change training methods, it specifically includes:1) network RPN pre-training is suggested in region: making RPN network carry out prison by ZF network using the training dataset of step 1 Superintend and direct pre-training；2) having for the convolutional neural networks of ZF model Fast R-CNN network pre-training: is carried out using the training dataset of step 1 Pre-training is supervised, the network parameter generated at the end of training using it is as the initiation parameter of joint training；3) RPN network fine tuning training: the positive sample of RPN network fine tuning training is alternative from generating with handmarking's frame and algorithm Frame registration is greater than 70% data sample, and negative sample, which is derived from, generates alternative frame, positive sample less than the algorithm of 30% registration Only indicate prospect, negative sample only indicates background；Operation is returned to carry out only for positive sample；It is abandoned when training all beyond image side The algorithm on boundary generates alternative frame, uses non-maxima suppression to the alternative frame collection generated beyond the algorithm behind boundary is removed；It is fixed Shared convolutional layer, that is, it is zero that its learning rate, which is arranged, is updated without parameter, only finely tunes the network parameter of the exclusive layer of RPN network, Realize shared convolution；4) Fast R-CNN network fine tuning training: Fast R-CNN network fine tuning training positive sample from handmarking's frame with Algorithm generates the data sample that alternative frame registration is greater than threshold region suggestion, and negative sample is raw from handmarking's frame and algorithm It is less than the data sample of threshold region suggestion at alternative frame registration, only finely tunes the exclusive layer of Fast R-CNN, is so far formed altogether Enjoy the Unified Network of convolution model.
- The point target detecting method 4. a kind of infrared defeated variation based on deep learning according to claim 3 is often generated heat, It is characterized in that, network RPN pre-training is suggested in the region, carries out end-to-end instruction especially by backpropagation and stochastic gradient descent Practice, trains this network according to " image-centric " sampling policy in FastR-CNN, each minimum step is by containing The single image of many positive negative samples forms, while using 0.01 mean value of standard deviation random to newly-increased layer for 0 Gaussian Profile Initialization.
- The point target detecting method 5. a kind of infrared defeated variation based on deep learning according to claim 3 is often generated heat, It is characterized in that, the threshold value is set as 0.4.
- The point target detecting method 6. a kind of infrared defeated variation based on deep learning according to claim 3 is often generated heat, It is characterized in that, training rate is 0.0001.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910612845.9A CN110334661A (en) | 2019-07-09 | 2019-07-09 | Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910612845.9A CN110334661A (en) | 2019-07-09 | 2019-07-09 | Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110334661A true CN110334661A (en) | 2019-10-15 |
Family
ID=68143898
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910612845.9A Pending CN110334661A (en) | 2019-07-09 | 2019-07-09 | Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110334661A (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110780164A (en) * | 2019-11-04 | 2020-02-11 | 华北电力大学（保定） | Insulator infrared fault positioning diagnosis method and device based on YOLO |
| CN110850723A (en) * | 2019-12-02 | 2020-02-28 | 西安科技大学 | Fault diagnosis and positioning method based on transformer substation inspection robot system |
| CN111046943A (en) * | 2019-12-09 | 2020-04-21 | 国网智能科技股份有限公司 | Method and system for automatically identifying state of isolation switch of transformer substation |
| CN111402253A (en) * | 2020-04-03 | 2020-07-10 | 华东交通大学 | Online monitoring method for state of power transmission and transformation equipment integrating edge calculation and deep learning |
| CN111398339A (en) * | 2020-03-26 | 2020-07-10 | 国网浙江省电力有限公司电力科学研究院 | Method and system for analyzing and judging heating defects of composite insulator of on-site overhead line |
| CN111444801A (en) * | 2020-03-18 | 2020-07-24 | 成都理工大学 | Real-time detection method for infrared target of unmanned aerial vehicle |
| CN111582477A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111738156A (en) * | 2020-06-23 | 2020-10-02 | 国网宁夏电力有限公司培训中心 | Intelligent inspection management method and system for state of high-voltage switchgear |
| CN111931058A (en) * | 2020-08-19 | 2020-11-13 | 中国科学院深圳先进技术研究院 | Sequence recommendation method and system based on adaptive network depth |
| CN111931593A (en) * | 2020-07-16 | 2020-11-13 | 上海无线电设备研究所 | Weak target detection method based on deep neural network and time-frequency image sequence |
| CN111948501A (en) * | 2020-08-05 | 2020-11-17 | 广州市赛皓达智能科技有限公司 | Automatic inspection equipment for power grid operation |
| CN111986425A (en) * | 2020-09-04 | 2020-11-24 | 江苏濠汉信息技术有限公司 | Power transmission channel early warning system based on infrared thermal imaging and early warning method thereof |
| CN112036464A (en) * | 2020-08-26 | 2020-12-04 | 国家电网有限公司 | Insulator infrared image fault detection method based on YOLOv3-tiny algorithm |
| CN112036463A (en) * | 2020-08-26 | 2020-12-04 | 国家电网有限公司 | Power equipment defect detection and identification method based on deep learning |
| CN112381784A (en) * | 2020-11-12 | 2021-02-19 | 国网浙江省电力有限公司信息通信分公司 | Equipment detecting system based on multispectral image |
| CN112396104A (en) * | 2020-11-17 | 2021-02-23 | 华中科技大学 | Plasma discharge identification method and system based on machine learning |
| CN112669316A (en) * | 2021-01-29 | 2021-04-16 | 南方电网调峰调频发电有限公司 | Power production abnormity monitoring method and device, computer equipment and storage medium |
| CN112733771A (en) * | 2021-01-18 | 2021-04-30 | 哈尔滨市科佳通用机电股份有限公司 | Railway train jumper wire foreign matter fault identification method and system |
| CN112734692A (en) * | 2020-12-17 | 2021-04-30 | 安徽继远软件有限公司 | Transformer equipment defect identification method and device |
| CN113011446A (en) * | 2019-12-20 | 2021-06-22 | 中国科学院沈阳自动化研究所 | Intelligent target identification method based on multi-source heterogeneous data learning |
| CN113033845A (en) * | 2021-04-25 | 2021-06-25 | 广东电网有限责任公司江门供电局 | Construction method and device for power transmission resource co-construction and sharing |
| CN113205178A (en) * | 2021-04-27 | 2021-08-03 | 特斯联科技集团有限公司 | Artificial intelligent infrared image sensing system and method |
| CN113269074A (en) * | 2021-05-19 | 2021-08-17 | 陕西科技大学 | Target detection method and security inspection robot |
| CN113342864A (en) * | 2021-06-02 | 2021-09-03 | 上海蓝色帛缔智能工程有限公司 | Method, system and device for monitoring heat productivity of electronic equipment |
| CN113469950A (en) * | 2021-06-08 | 2021-10-01 | 海南电网有限责任公司电力科学研究院 | Method for diagnosing abnormal heating defect of composite insulator based on deep learning |
| CN113673579A (en) * | 2021-07-27 | 2021-11-19 | 国网湖北省电力有限公司营销服务中心(计量中心) | Power load classification algorithm based on small samples |
| CN117314896A (en) * | 2023-11-28 | 2023-12-29 | 国网湖北省电力有限公司 | Power system abnormality detection method and system based on deep learning |
| CN117538658A (en) * | 2023-11-16 | 2024-02-09 | 深圳市美信检测技术股份有限公司 | Artificial intelligence fault positioning method and device based on infrared spectrum and thermal imaging |
| CN117911401A (en) * | 2024-03-15 | 2024-04-19 | 国网山东省电力公司泗水县供电公司 | Power equipment fault detection method, system, storage medium and equipment |
-
2019
- 2019-07-09 CN CN201910612845.9A patent/CN110334661A/en active Pending
Cited By (39)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110780164A (en) * | 2019-11-04 | 2020-02-11 | 华北电力大学（保定） | Insulator infrared fault positioning diagnosis method and device based on YOLO |
| CN110780164B (en) * | 2019-11-04 | 2022-03-25 | 华北电力大学（保定） | Insulator infrared fault positioning diagnosis method and device based on YOLO |
| CN110850723A (en) * | 2019-12-02 | 2020-02-28 | 西安科技大学 | Fault diagnosis and positioning method based on transformer substation inspection robot system |
| CN111046943A (en) * | 2019-12-09 | 2020-04-21 | 国网智能科技股份有限公司 | Method and system for automatically identifying state of isolation switch of transformer substation |
| CN113011446A (en) * | 2019-12-20 | 2021-06-22 | 中国科学院沈阳自动化研究所 | Intelligent target identification method based on multi-source heterogeneous data learning |
| CN113011446B (en) * | 2019-12-20 | 2023-08-04 | 中国科学院沈阳自动化研究所 | Intelligent target recognition method based on multi-source heterogeneous data learning |
| CN111444801A (en) * | 2020-03-18 | 2020-07-24 | 成都理工大学 | Real-time detection method for infrared target of unmanned aerial vehicle |
| CN111398339A (en) * | 2020-03-26 | 2020-07-10 | 国网浙江省电力有限公司电力科学研究院 | Method and system for analyzing and judging heating defects of composite insulator of on-site overhead line |
| CN111398339B (en) * | 2020-03-26 | 2022-08-30 | 国网浙江省电力有限公司电力科学研究院 | Method and system for analyzing and judging heating defects of composite insulator of on-site overhead line |
| CN111402253A (en) * | 2020-04-03 | 2020-07-10 | 华东交通大学 | Online monitoring method for state of power transmission and transformation equipment integrating edge calculation and deep learning |
| CN111582477B (en) * | 2020-05-09 | 2023-08-29 | 北京百度网讯科技有限公司 | Training method and device for neural network model |
| CN111582477A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Training method and device of neural network model |
| CN111738156A (en) * | 2020-06-23 | 2020-10-02 | 国网宁夏电力有限公司培训中心 | Intelligent inspection management method and system for state of high-voltage switchgear |
| CN111931593A (en) * | 2020-07-16 | 2020-11-13 | 上海无线电设备研究所 | Weak target detection method based on deep neural network and time-frequency image sequence |
| CN111931593B (en) * | 2020-07-16 | 2024-04-26 | 上海无线电设备研究所 | Weak target detection method based on deep neural network and time-frequency image sequence |
| CN111948501A (en) * | 2020-08-05 | 2020-11-17 | 广州市赛皓达智能科技有限公司 | Automatic inspection equipment for power grid operation |
| CN111948501B (en) * | 2020-08-05 | 2022-12-13 | 广州市赛皓达智能科技有限公司 | Automatic inspection equipment for power grid operation |
| CN111931058A (en) * | 2020-08-19 | 2020-11-13 | 中国科学院深圳先进技术研究院 | Sequence recommendation method and system based on adaptive network depth |
| CN111931058B (en) * | 2020-08-19 | 2024-01-05 | 中国科学院深圳先进技术研究院 | Sequence recommendation method and system based on self-adaptive network depth |
| CN112036464A (en) * | 2020-08-26 | 2020-12-04 | 国家电网有限公司 | Insulator infrared image fault detection method based on YOLOv3-tiny algorithm |
| CN112036463A (en) * | 2020-08-26 | 2020-12-04 | 国家电网有限公司 | Power equipment defect detection and identification method based on deep learning |
| CN111986425A (en) * | 2020-09-04 | 2020-11-24 | 江苏濠汉信息技术有限公司 | Power transmission channel early warning system based on infrared thermal imaging and early warning method thereof |
| CN112381784A (en) * | 2020-11-12 | 2021-02-19 | 国网浙江省电力有限公司信息通信分公司 | Equipment detecting system based on multispectral image |
| CN112396104A (en) * | 2020-11-17 | 2021-02-23 | 华中科技大学 | Plasma discharge identification method and system based on machine learning |
| CN112734692B (en) * | 2020-12-17 | 2023-12-22 | 国网信息通信产业集团有限公司 | Defect identification method and device for power transformation equipment |
| CN112734692A (en) * | 2020-12-17 | 2021-04-30 | 安徽继远软件有限公司 | Transformer equipment defect identification method and device |
| CN112733771A (en) * | 2021-01-18 | 2021-04-30 | 哈尔滨市科佳通用机电股份有限公司 | Railway train jumper wire foreign matter fault identification method and system |
| CN112669316A (en) * | 2021-01-29 | 2021-04-16 | 南方电网调峰调频发电有限公司 | Power production abnormity monitoring method and device, computer equipment and storage medium |
| CN113033845A (en) * | 2021-04-25 | 2021-06-25 | 广东电网有限责任公司江门供电局 | Construction method and device for power transmission resource co-construction and sharing |
| CN113205178A (en) * | 2021-04-27 | 2021-08-03 | 特斯联科技集团有限公司 | Artificial intelligent infrared image sensing system and method |
| CN113269074A (en) * | 2021-05-19 | 2021-08-17 | 陕西科技大学 | Target detection method and security inspection robot |
| CN113342864A (en) * | 2021-06-02 | 2021-09-03 | 上海蓝色帛缔智能工程有限公司 | Method, system and device for monitoring heat productivity of electronic equipment |
| CN113469950A (en) * | 2021-06-08 | 2021-10-01 | 海南电网有限责任公司电力科学研究院 | Method for diagnosing abnormal heating defect of composite insulator based on deep learning |
| CN113673579A (en) * | 2021-07-27 | 2021-11-19 | 国网湖北省电力有限公司营销服务中心(计量中心) | Power load classification algorithm based on small samples |
| CN113673579B (en) * | 2021-07-27 | 2024-05-28 | 国网湖北省电力有限公司营销服务中心(计量中心) | Small sample-based electricity load classification algorithm |
| CN117538658A (en) * | 2023-11-16 | 2024-02-09 | 深圳市美信检测技术股份有限公司 | Artificial intelligence fault positioning method and device based on infrared spectrum and thermal imaging |
| CN117314896A (en) * | 2023-11-28 | 2023-12-29 | 国网湖北省电力有限公司 | Power system abnormality detection method and system based on deep learning |
| CN117314896B (en) * | 2023-11-28 | 2024-02-06 | 国网湖北省电力有限公司 | Power system abnormality detection method and system based on deep learning |
| CN117911401A (en) * | 2024-03-15 | 2024-04-19 | 国网山东省电力公司泗水县供电公司 | Power equipment fault detection method, system, storage medium and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110334661A (en) | Infrared power transmission and transformation abnormal heating point target detecting method based on deep learning | |
| CN110598736B (en) | Power equipment infrared image fault positioning, identifying and predicting method | |
| Wang et al. | Machine vision for natural gas methane emissions detection using an infrared camera | |
| Nasiri et al. | Intelligent fault diagnosis of cooling radiator based on deep learning analysis of infrared thermal images | |
| CN108985376B (en) | Rotary kiln sequence working condition identification method based on convolution-cyclic neural network | |
| CN106919978B (en) | Method for identifying and detecting parts of high-speed rail contact net supporting device | |
| CN111325347B (en) | Automatic danger early warning description generation method based on interpretable visual reasoning model | |
| CN109101906A (en) | A kind of converting station electric power equipment infrared image exception real-time detection method and device | |
| CN110378408A (en) | Power equipment image-recognizing method and device based on transfer learning and neural network | |
| CN110017901A (en) | The infared spectrum diagnostic method of electric equipment operation state in a kind of electric system | |
| CN114069838A (en) | Transformer substation robot intelligent inspection system and method with intelligent sensor actively cooperated | |
| CN111209832B (en) | Auxiliary obstacle avoidance training method, equipment and medium for substation inspection robot | |
| CN109919032A (en) | A kind of video anomaly detection method based on action prediction | |
| CN111695731A (en) | Load prediction method, system and equipment based on multi-source data and hybrid neural network | |
| Song et al. | Deformable YOLOX: Detection and rust warning method of transmission line connection fittings based on image processing technology | |
| Mlakić et al. | Deep learning method and infrared imaging as a tool for transformer faults detection | |
| Zhang et al. | Spatial-temporal dual-channel adaptive graph convolutional network for remaining useful life prediction with multi-sensor information fusion | |
| CN110378407A (en) | Power equipment image-recognizing method and device based on parametric texture and neural network | |
| CN113361520B (en) | Transmission line equipment defect detection method based on sample offset network | |
| Lu et al. | Deep learning based fusion of RGB and infrared images for the detection of abnormal condition of fused magnesium furnace | |
| Anochi et al. | New learning strategy for supervised neural network: MPCA meta-heuristic approach | |
| CN115995051A (en) | Substation equipment fault period identification method based on minimum residual error square sum method | |
| Zhang | Computer image processing and neural network technology for thermal energy diagnosis of boiler plants | |
| Zhang et al. | Research on intelligent health management technology of opto-electronic equipment | |
| Wang et al. | Design and Research of Smart Grid Based on Artificial Intelligence |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191015 |
|
| WD01 | Invention patent application deemed withdrawn after publication |