The content of the invention
The invention provides a kind of power equipment thermal fault detection method, system and electronic equipment, it is intended at least certain
Solves one of above-mentioned technical problem of the prior art in degree.
In order to solve the above problems, the invention provides following technical scheme:
A kind of power equipment thermal fault detection method, including:
Step a:The infrared image of power equipment is gathered, convolutional neural networks model is built according to the infrared image;
Step b:Infrared image to be detected is inputted into the convolutional neural networks model, passes through the convolutional neural networks
The temperature scale and power equipment that Model Identification goes out in infrared image;
Step c:According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature
Degree extracts the rgb value of identified power equipment with reference to table, by the rgb value of extraction and the rgb value with temperature with reference in table
Rgb value be compared, obtain the temperature results of identified power equipment;
Step d:Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judging should
Whether power equipment there is hot stall.
The technical scheme that the embodiment of the present invention is taken also includes:In the step a, described built according to infrared image is rolled up
Product neural network model also includes:Training sample is made according to the infrared image of collection, convolution is built according to the training sample
Neural network model.
The technical scheme that the embodiment of the present invention is taken also includes:It is described that training sample tool is made according to the infrared image of collection
Body includes:Four apex coordinates of rectangle frame where marking the power equipment in every width infrared image respectively, according to power equipment
Type set corresponds to the label of rectangle frame, according to four apex coordinates and rectangle of rectangle frame where picture name, power equipment
Box label generates the labeled data of every width infrared image.
The technical scheme that the embodiment of the present invention is taken also includes:In the step a, the convolutional neural networks model bag
RPN convolutional neural networks and Fast-RCNN convolutional neural networks are included, the RPN convolutional neural networks are used to handle training sample,
Obtain the object candidate area in infrared image;Input of the object candidate area as Fast-RCNN convolutional neural networks,
The temperature scale and power equipment in credible degree identification object candidate area exported according to Fast-RCNN convolutional neural networks.
The technical scheme that the embodiment of the present invention is taken also includes:Also include after the step a:To the convolutional Neural net
Network model is trained;The training method that is trained to convolutional neural networks model is specially:
Step a1:The training parameter in the RPN convolutional neural networks is initialized with small random number;
Step a2:Training sample is input in the RPN convolutional neural networks after initialization, carries out propagated forward, is calculated
Differentiate classification and the Softmax penalty values of rectangle box label, calculate rectangle frame where object candidate area parameter and power equipment
SmoothL1 penalty values;
Step a3:Using stochastic gradient descent algorithm, each layer of cost function ladder in RPN convolutional neural networks is calculated respectively
Angle value, update the training parameter in RPN convolutional neural networks with Grad;
Step a4:Using backpropagation BP algorithm, each layer in RPN convolutional neural networks of error-sensitivity is calculated, is used
Error-sensitivity adjusts the training parameter of RPN convolutional neural networks, makes Softmax penalty values and SmoothL1 penalty values minimum,
The RPN convolutional neural networks trained;
Step a5:The RPN convolutional neural networks that training sample input is trained, obtain initial target candidate region;
Step a6:Fast-RCNN convolutional neural networks are instructed with the output of RPN convolutional neural networks and sample data
Practice, the Fast-RCNN convolutional neural networks trained, and obtain the network parameter of five layers of shared convolutional layer;
Step a7:After RPN convolutional neural networks being reinitialized using the network parameter of five layers of shared convolutional layer, input instruction
Practice sample re -training RPN convolutional neural networks, and export new object candidate area;
Step a8:Fast-RCNN convolutional Neural nets are reinitialized using the network parameter of five layers of shared convolutional layer
Network, and new object candidate area and sample data re -training Fast-RCNN convolutional neural networks are used, trained
The network parameter of Fast-RCNN convolutional neural networks.
The technical scheme that the embodiment of the present invention is taken also includes:It is described to obtain identified electric power and set in the step c
Standby temperature results specifically include:
Step c1:It is determined that environment temperature during collection infrared image, and determine the temperature scale in infrared image;
Step c2:Temperature maximum region and temperature minimum value region are distinguished according to the position of temperature scale, read
Numeral in temperature maximum region and temperature minimum value region, obtain temperature maximum and temperature minimum in the infrared image
Value;
Step c3:The value matrixs of tri- values of temperature scale R, G, B are calculated, in the centre position phase selection same column of three matrixes,
Form a rgb value matrix;
Step c4:Each row medium temperature of R, G, B is determined according to the border rgb value of the graded of R, G, B value and temperature scale
The bound of anale settting scale, and remove the redundance up and down of the rgb value matrix, obtain new rgb value matrix;
Step c5:According to new rgb value matrix and temperature maximum and temperature minimum value generation rgb value and temperature reference
Table;
Step c6:Maximum temperature of the rgb value maximum pixel point of power equipment as power equipment is extracted, by power equipment
Maximum temperature and rgb value be compared with temperature with reference to the rgb value in table, obtain the temperature results of the power equipment.
Another technical scheme that the embodiment of the present invention is taken is:A kind of power equipment thermal fault detection system, including:
Image capture module:For gathering the infrared image of power equipment;
Model construction module:For building convolutional neural networks model according to the infrared image;
Target identification module:For infrared image to be detected to be inputted into the convolutional neural networks model, by described
The temperature scale and power equipment that convolutional neural networks Model Identification goes out in infrared image;
Temperature detecting module:Given birth to for the rgb value according to the temperature scale pixel identified and temperature scale bound
Into rgb value and temperature with reference to table, and extract the rgb value of identified power equipment, by the rgb value of extraction and the rgb value with
Temperature is compared with reference to the rgb value in table, obtains the temperature results of identified power equipment;
Fault diagnosis module:For being carried out according to the temperature results of power equipment of the network system diagnostic criteria to being identified
Diagnosis, judges whether the power equipment hot stall occurs.
The technical scheme that the embodiment of the present invention is taken also includes sample and makes module, and the sample makes module and is used for basis
The infrared image of collection makes training sample, and the model construction module builds convolutional neural networks model according to training sample.
The technical scheme that the embodiment of the present invention is taken also includes:The sample makes module making training sample:
Four apex coordinates of rectangle frame where marking the power equipment in every width infrared image respectively, according to power equipment type set
The label of corresponding rectangle frame, given birth to according to four apex coordinates of rectangle frame where picture name, power equipment and rectangle box label
Into the labeled data of every width infrared image.
The technical scheme that the embodiment of the present invention is taken also includes:The convolutional neural networks model includes RPN convolutional Neurals
Network and Fast-RCNN convolutional neural networks, the RPN convolutional neural networks are used to handle training sample, obtain infrared image
In object candidate area;Input of the object candidate area as Fast-RCNN convolutional neural networks, according to Fast-
Temperature scale and power equipment in the credible degree identification object candidate area of RCNN convolutional neural networks output.
The technical scheme that the embodiment of the present invention is taken also includes model training module, and the model training module is used for institute
Convolutional neural networks model is stated to be trained;Described be trained to convolutional neural networks model specifically includes:With small random number
Initialize the training parameter in the RPN convolutional neural networks;Training sample is input to the RPN convolutional Neural nets after initialization
In network, the Softmax penalty values of propagated forward, computational discrimination classification and rectangle box label are carried out, calculate object candidate area ginseng
The SmoothL1 penalty values of rectangle frame where number and power equipment;Using stochastic gradient descent algorithm, RPN convolution god is calculated respectively
Through each layer of cost function Grad in network, the training parameter in RPN convolutional neural networks is updated with Grad;Using reverse
BP algorithm is propagated, calculates each layer in RPN convolutional neural networks of error-sensitivity, RPN convolution god is adjusted with error-sensitivity
Training parameter through network, make Softmax penalty values and SmoothL1 penalty values minimum, the RPN convolutional Neurals trained
Network;The RPN convolutional neural networks that training sample input is trained, obtain initial target candidate region;With RPN convolutional Neurals
The output of network and sample data are trained to Fast-RCNN convolutional neural networks, the Fast-RCNN convolution trained
Neutral net, and obtain the network parameter of five layers of shared convolutional layer;It is again initial using the network parameter of five layers of shared convolutional layer
After changing RPN convolutional neural networks, training sample re -training RPN convolutional neural networks are inputted, and export new target candidate area
Domain;Fast-RCNN convolutional neural networks are reinitialized using the network parameter of five layers of shared convolutional layer, and are used newly
Object candidate area and sample data re -training Fast-RCNN convolutional neural networks, the Fast-RCNN convolution trained
The network parameter of neutral net.
The technical scheme that the embodiment of the present invention is taken also includes:The temperature detecting module obtains identified power equipment
Temperature results specifically include:It is determined that environment temperature during collection infrared image, and determine the temperature scale in infrared image;Root
Distinguish temperature maximum region and temperature minimum value region according to the position of temperature scale, read temperature maximum region respectively
With the numeral in temperature minimum value region, the temperature maximum and temperature minimum value in the infrared image are obtained;Calculate the scale of thermometer
The value matrix of tri- values of chi R, G, B, in the centre position phase selection same column of three matrixes, form a rgb value matrix;According to R, G,
The graded of B values and the border rgb value of temperature scale determine the bound of temperature scale in each row of R, G, B, and remove institute
The redundance up and down of rgb value matrix is stated, obtains new rgb value matrix;According to new rgb value matrix and temperature maximum and
Temperature minimum value generates rgb value with temperature with reference to table;Extract power equipment rgb value maximum pixel point as power equipment most
High-temperature, the maximum temperature of power equipment and rgb value are compared with temperature with reference to the rgb value in table, obtain the electric power
The temperature results of equipment.
The another technical scheme that the embodiment of the present invention is taken is:A kind of electronic equipment, including:
At least one processor;And
The memory being connected with least one processor communication;Wherein,
The memory storage has can be by the instruction of one computing device, and the instruction is by least one place
Manage device to perform, so that at least one processor is able to carry out the following operation of above-mentioned power equipment thermal fault detection method:
The infrared image of power equipment is gathered, convolutional neural networks model is built according to the infrared image;
Infrared image to be detected is inputted into the convolutional neural networks model, known by the convolutional neural networks model
The temperature scale and power equipment not gone out in infrared image;
According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature reference
Table, and the rgb value of identified power equipment is extracted, by the rgb value of extraction and the rgb value with temperature with reference to the RGB in table
Value is compared, and obtains the temperature results of identified power equipment;
Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judge that the electric power is set
It is standby whether hot stall occur.
Relative to prior art, beneficial effect caused by the embodiment of the present invention is:The power equipment of the embodiment of the present invention
Thermal fault detection method, system and electronic equipment go out a special target identification convolution using Faster-RCNN Algorithm for Training
Neural network model, the model can identify power equipment and temperature scale in image by the infrared image of input, and make
The temperature of power equipment is read with rgb value Comparison Method, eventually through《DL/T 664-2008 charging equipments infrared diagnostics application rule
Model》In standard judge whether power equipment hot stall occurs.The present invention efficiently, is accurately known by convolutional neural networks model
Other power equipment, temperature is accurately read by rgb value, the hot stall of final efficient, reliable diagnosing electric power.Present invention leather
The new mode of power equipment thermal fault detection, lift the intelligent level of network system.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
The concept of deep learning comes from artificial neural network, and its Research Significance is in the nerve net in establishing, simulate human brain
Network, the mechanism for imitating human brain is used for understanding and handling various data, for analyzing and learning.Deep learning is by combining low layer spy
Sign forms more abstract high-rise expression attribute classification or feature, to find that the distributed nature of data represents.Convolutional Neural net
Network model is one kind of deep learning method, and the special construction that convolutional neural networks model is shared with its local weight is at image
There is the superiority of uniqueness in terms of reason, it is laid out the biological neural network closer to reality, and weights are shared to reduce network
Complexity, the images of particularly more dimensional input vectors can directly input network this feature and avoid feature extraction and sorted
The complexity of data reconstruction in journey.
The power equipment thermal fault detection method of the embodiment of the present invention for needed to use in network system infrared imaging come
Read power equipment temperature and then detect the demand of failure, going out a special target using Faster-RCNN Algorithm for Training knows
Other convolutional neural networks model, the model can identify power equipment and the scale of thermometer in image by the infrared image of input
Chi, and the temperature of rgb value Comparison Method reading power equipment is used, eventually through《DL/T 664-2008 charging equipment infrared diagnosticses
Using specification》In standard judge whether power equipment hot stall occurs.
Referring to Fig. 1, it is the overall flow figure of the power equipment thermal fault detection method of the embodiment of the present invention.It is of the invention real
The power equipment thermal fault detection method for applying example comprises the following steps:
Step a:The infrared image of power equipment is gathered, convolutional neural networks model is built according to infrared image;
Step b:Infrared image to be detected is inputted into convolutional neural networks model, gone out by convolutional neural networks Model Identification
Temperature scale and power equipment in infrared image to be detected;
Step c:According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature
Degree extracts the rgb value of identified power equipment with reference to table, by the rgb value of extraction and the rgb value with temperature with reference in table
Rgb value be compared, obtain the temperature results of identified power equipment;
Step d:Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judging should
Whether power equipment there is hot stall.
Specifically, the power equipment thermal fault detection method of the embodiment of the present invention includes target identification stage and temperature detection
In the stage, target identification stage and temperature detection stage are specifically described individually below.
Referring to Fig. 2, it is the target identification flow chart of the embodiment of the present invention.The target identification method bag of the embodiment of the present invention
Include following steps:
Step 100:Gather the infrared image of power equipment;
In step 100, the infrared image of all types of power equipments, and the electricity per class are included in the infrared image of collection
The infrared image acquisition quantity of power equipment can be set according to needed for deep learning network training.
Step 110:Infrared image based on collection, it is used as training by making the infrared picture library with mark and label
Sample;
In step 110, the production method of training sample is:All kinds of electric power are chosen respectively from the infrared image of collection to set
Standby a number of infrared image is labeled, and specific notation methods are:According to VOC2007 forms, OPENCV images are used
Handling implement, four apex coordinates of rectangle frame where marking the power equipment in every width infrared image respectively, and according to electric power
The rectangle box label of the corresponding rectangle frame of device type setting.I.e.:Every width infrared image includes 6 labeled data respectively:Picture name,
Four apex coordinates of rectangle frame where rectangle box label, equipment, and 6 labeled data are generated into txt file, use Matlab
Txt file is changed into xml document, (mark of 6 labeled data for determining particular location in each image is used for network training
It is what to gaze at mark).
Step 120:Training sample based on making, the convolutional Neural of target identification is built using Faster-RCNN algorithms
Network model;
In the step 120, the present invention uses the more advanced Faster-RCNN algorithms of current detection accuracy and speed,
A RPN convolutional neural networks and a Fast-RCNN convolutional neural networks are respectively trained, the RPN convolutional Neural nets trained
Network is responsible for handling sample data, obtains the object candidate area in infrared image;The object candidate area is as Fast-RCNN volumes
The input of product neutral net, the mesh in credible degree identification object candidate area exported according to Fast-RCNN convolutional neural networks
Mark.
In above-mentioned, RPN convolutional neural networks are made up of six convolutional layers, two full articulamentums and a Softmax layer.Six
In individual convolutional layer, first five layer is characterized extract layer, and layer 6 is characterized mapping layer.Layer 7 and the 8th layer are full connection at the same level
Layer, layer 7 are responsible for classification, obtain the object candidate area in infrared image;8th layer of responsible recurrence, output target candidate area
The confidence level in domain.Adjustment is normalized to the confidence level of the 8th layer of output in Softmax layers.Finally according to object candidate area
Confidence level exports the location parameter of reliable object candidate area.
Fast-RCNN convolutional neural networks are by five convolutional layers, a ROI ponds layer, several full articulamentums and one
Softmax layers are formed.According to Faster-RCNN algorithmic characteristics, Fast-RCNN five convolutional layers are as five spies with RPN
It is shared convolutional layer to levy extract layer.Input of the output of convolution results and RPN after five layers of convolution as ROI ponds layer, ROI ponds
It is that several full articulamentums do nonlinear transformation to feature after change layer, two full articulamentums of peer is reused afterwards, with RPN convolution god
Through network, the location parameter of classification layer output object candidate area, the confidence level that layer exports object candidate area is returned.
Softmax layers confidence level are normalized adjustment, and the location parameter of combining target candidate region obtains object candidate area, make
For the output of whole Fast-RCNN convolutional neural networks.
Convolutional neural networks model structure is as shown in Figure 3.In figure 3, Conv is convolutional layer;ReLu(Rectified
Linear unit) it is activation primitive;Pooling operates for pondization, and feature maps are characterized mapping;Reg layer are side
The full articulamentum that frame returns;Cls layer are the full articulamentum of classification;Softmax is regression function;Proposals is candidate
Frame;ROI Pooling are the pondization operation for ROIs, and ROI is writing a Chinese character in simplified form for Region of interest, refers to Faste-
In RCNN structures, after rpn layers, box frames corresponding to caused proposal;FC is full articulamentum, and FCs is that multilayer connects entirely
Layer.
In the embodiment of the present invention, the building mode of convolutional neural networks model specifically includes following steps:
Step 121:Build five layers of convolutional neural networks:First layer uses 96 convolution kernels, and convolution kernel window size is 7 ×
7 pixels, convolution kernel window are that interval carries out convolution operation over an input image with 2, export 96 characteristic patterns, each characteristic pattern
Dimensionality reduction is carried out by a down-sampling, connection ReLU non-linear layers do nonlinear transformation, reuse core window size as 3 × 3
Pixel, down-sampling is carried out at intervals of 2 maximum pond layer, the characteristic pattern after dimensionality reduction is input to the second layer.The second layer uses 256
Individual convolution kernel, convolution kernel window size are 5 × 5 pixels, and convolution kernel window does convolution with 2 to be spaced in the output of first layer
Operation, 256 characteristic patterns are exported, each characteristic pattern carries out dimensionality reduction by a down-sampling, and connection ReLU non-linear layers do non-thread
Property conversion, it be 2 × 2 pixels to reuse core window size, at intervals of 2 maximum pond layer progress down-sampling, after dimensionality reduction
Characteristic pattern is input to third layer.384 third layer, the 4th layer of use convolution kernels, convolution kernel window size are 3 × 3 pixels, volume
Product core window carries out convolution operation with 1 to be spaced in the output of respective preceding layer, exports 384 characteristic patterns, each characteristic pattern
Connection ReLU non-linear layers do nonlinear transformation, and the characteristic pattern after conversion is input into next layer.Layer 5 uses 256 convolution
Core, convolution kernel window size are 3 × 3 pixels, and convolution kernel window is that interval carries out convolution operation over an input image with 1, defeated
Go out 256 characteristic patterns, each characteristic pattern connection ReLU non-linear layers do nonlinear transformation, the characteristic pattern after conversion is input to down
One layer.Five layers of convolutional neural networks of the above are RPN and Fast-RCNN shared convolutional layer.
Step 122:Build RPN convolutional neural networks;After building five layers of convolutional neural networks, RPN convolutional neural networks
Layer 6 convolutional layer need to be built as feature extraction layer.Layer 6 structure is consistent with layer 5, exports 256 characteristic patterns, feature
Figure is input to layer 7 and the 8th layer after doing nonlinear transformation.Reference frame is built, for each position of infrared image, is considered
To 9 possible windows:Three kinds of areas { 1282,2562,5122 }, three kinds of ratios { 1:1,1:2,2:1}.Layer 7 and the 8th layer
For the full articulamentum of peer, layer 7 is the recurrence layer of 24 neurons, the 8th layer be 48 neurons classification layer.Layer 7
A Softmax layer is reused below.
Step 123:Build Fast-RCNN convolutional neural networks;After building five layers of convolutional neural networks, Fast-RCNN
The target candidate area exported using a ROI ponds layer, ROI floor with 256 characteristic patterns and RPN convolutional neural networks of layer 5
Domain is input, and object candidate area is mapped on characteristic pattern, forms the characteristic pattern output of 6 × 6 pixels.After the layer of ROI ponds,
Nonlinear transformation is done to feature using several full articulamentums, the neuronal quantity of full articulamentum is 4096.In nonlinear transformation
Afterwards, there is the full articulamentum classification layer of peer and return layer, classification layer is the full connection of 8 neurons, and recurrence layer is 2 nerves
The full connection of member, return layer and reuse a Softmax layer.
Step 130:Convolutional neural networks model is trained;
It is specific as shown in figure 4, model training flow chart for the embodiment of the present invention.The model training side of the embodiment of the present invention
Method comprises the following steps:
Step 131:RPN convolutional neural networks are initialized, RPN convolution god is initialized with different small random numbers (0~1)
Through the training parameter in network;
Step 132:Training sample is input in RPN convolutional neural networks, carry out propagated forward, computational discrimination classification and
The Softmax penalty values of rectangle box label, the SmoothL1 losses of rectangle frame where calculating object candidate area parameter and equipment
Value;Wherein, differentiate that classification refers to:Can be according to the characteristic value of current label target (training sample) after propagated forward, and read
The characteristic value of picture is taken, judges the probability of the possible target type of current label target.
Step 133:Using stochastic gradient descent algorithm, each layer of cost function in RPN convolutional neural networks is calculated respectively
Grad, update the training parameter in RPN convolutional neural networks with Grad;
Step 134:Using backpropagation BP algorithm, each layer in RPN convolutional neural networks of error-sensitivity is calculated, is used
Error-sensitivity adjusts the training parameter of RPN convolutional neural networks, makes Softmax penalty values and SmoothL1 penalty values minimum,
The RPN convolutional neural networks trained;
Step 135:The RPN convolutional neural networks that training sample input is trained, obtain initial target candidate region;
Step 136:Using the initialization same with RPN convolutional neural networks and training method, to Fast-RCNN convolution god
It is trained through network, inputs the output for RPN convolutional neural networks and sample data, Fast-RCNN volumes trained
Product neutral net, and obtain the network parameter of five layers of shared convolutional layer;
Step 137:After RPN convolutional neural networks being reinitialized using the network parameter of five layers of shared convolutional layer, input
Training sample re -training RPN convolutional neural networks, and export new object candidate area;
Step 138:Fast-RCNN convolutional neural networks are reinitialized using the network parameter of five layers of shared convolutional layer,
And the new object candidate area and sample data exported using in step 137 is used as input, re -training Fast-RCNN convolution
Neutral net, each layer network parameter of the Fast-RCNN convolutional neural networks trained.
Step 140:Infrared image to be measured is inputted into convolutional neural networks model, the result accuracy rate of test model.
Step 150:Infrared image input to be detected is trained and by the convolutional neural networks model of test, passed through
The power equipment and temperature scale that convolutional neural networks Model Identification goes out in infrared image, and and mark out power equipment mesh respectively
Mark region and temperature scale target area.
Referring to Fig. 5, it is the temperature detection flow chart of the embodiment of the present invention.The temperature checking method bag of the embodiment of the present invention
Include following steps:
Step 200:It is determined that environment temperature during collection infrared image, and determine the temperature scale in infrared image;
Step 210:Temperature maximum region and temperature minimum value are distinguished according to the position of temperature scale target area
Region, the numeral in temperature maximum region and temperature minimum value region is read by optical recognition program tesseract respectively,
Obtain the temperature maximum Tmax and temperature minimum value Tmin in the infrared image;
Step 220:The value matrix of tri- values of R, G, B in temperature scale target area is calculated, in the interposition of three matrixes
Phase selection same column is put, forms a rgb value matrix, the bound rgb value of temperature scale is all identical in every width infrared image;
Step 230:Because temperature scale target area can not be accurately positioned the bound of temperature scale, according to R, G, B value
Graded and temperature scale border rgb value determine R, G, B respectively row in temperature scale bound, remove rgb value square
The redundance up and down of battle array, obtains new rgb value matrix;
Step 240:Joined according to new rgb value matrix and temperature maximum and temperature minimum value generation rgb value with temperature
According to table;
Step 250:Maximum temperature of the rgb value maximum pixel point in power equipment callout box as power equipment is extracted,
The maximum temperature of power equipment and rgb value are compared with temperature with reference to the rgb value in table, obtain the temperature knot of power equipment
Fruit;
In step 250, in power equipment target area, because the heating temp of power equipment can be higher than environment temperature
Degree, it is more obvious the power equipment of hot stall to be present, and therefore, the rgb value peak that can extract in power equipment target area is made
Point is read for the temperature of the power equipment, is contrasted using the rgb value and rgb value of the point with temperature with reference to the rgb value in table, you can
Obtain the temperature results of the power equipment.
Step 260:According to device type, use《DL/T 664-2008 charging equipment infrared diagnostics application specifications》In
Standard diagnoses to the temperature results of power equipment, judges whether power equipment hot stall occurs.
In step 260, the maximum temperature T of power equipment is obtained according to temperature detection, electric power is obtained according to device type
Environment temperature when normal temperature Tn corresponding to equipment, Te are collection infrared image, is provided in infrared image.Thus electricity is obtained
The temperature rise value Tr=T-Te of power equipment, the temperature approach Td=T-Tn of power equipment, relative temperature difference value δ=(T- of power equipment
Tn)/(T-Te).To the temperature rise value of power equipment, temperature approach and relative temperature difference value, use《DL/T 664-2008 charging equipments are red
Outer diagnostic application specification》In standard diagnosed, from judging whether the power equipment hot stall occurs.
Referring to Fig. 6, it is the structural representation of the power equipment thermal fault detection system of the embodiment of the present invention.It is of the invention real
Applying the power equipment thermal fault detection system of example includes image capture module, sample making module, model construction module, model instruction
Practice module, model measurement module, target identification module, temperature detecting module and fault diagnosis module.
Image capture module:For gathering the infrared image of power equipment;Wherein, comprising all in the infrared image of collection
The infrared image of the power equipment of type, and can be according to deep learning network per the infrared image acquisition quantity of class power equipment
Set needed for training.
Sample makes module:For the infrared image based on collection, by making the infrared picture with mark and label
Storehouse is as training sample;Wherein, the production method of training sample is:All kinds of electric power are chosen respectively from the infrared image of collection to set
Standby a number of infrared image is labeled, and specific notation methods are:According to VOC2007 forms, OPENCV images are used
Handling implement, four apex coordinates of rectangle frame where marking the power equipment in every width infrared image respectively, and according to electric power
The label of the corresponding rectangle frame of device type setting.I.e.:Every width infrared image includes 6 labeled data respectively:Picture name, rectangle frame
Four apex coordinates of label, rectangle frame where equipment, and 6 labeled data are generated into txt files, using Matlab by txt
File changes into xml document, is used for network training.
Model construction module:For the training sample based on making, target identification is built using Faster-RCNN algorithms
Convolutional neural networks model;Wherein, the present invention is calculated using the more advanced Faster-RCNN of current detection accuracy and speed
Method, a RPN convolutional neural networks and a Fast-RCNN convolutional neural networks are respectively trained, the RPN convolutional Neurals trained
Network is responsible for handling sample data, obtains the object candidate area in infrared image;The object candidate area is as Fast-RCNN
The input of convolutional neural networks, according in the credible degree identification object candidate area of Fast-RCNN convolutional neural networks output
Target.
Specifically, model construction module includes shared convolutional layer construction unit, RPN network structions unit and Fast-RCNN
Network struction unit;
Shared convolutional layer construction unit:For building five layers of convolutional neural networks:First layer uses 96 convolution kernels, convolution
Core window size is 7 × 7 pixels, and convolution kernel window is that interval carries out convolution operation over an input image with 2, exports 96 spies
Sign figure, each characteristic pattern carry out dimensionality reduction by a down-sampling, and connection ReLU non-linear layers do nonlinear transformation, reuse core window
Mouth size is 3 × 3 pixels, carries out down-sampling at intervals of 2 maximum pond layer, the characteristic pattern after dimensionality reduction is input into second
Layer.The second layer uses 256 convolution kernels, and convolution kernel window size is 5 × 5 pixels, and convolution kernel window is to be spaced in first with 2
Convolution operation is done in the output of layer, exports 256 characteristic patterns, each characteristic pattern carries out dimensionality reduction, connection by a down-sampling
ReLU non-linear layers do nonlinear transformation, and it is 2 × 2 pixels to reuse core window size, are carried out at intervals of 2 maximum pond layer
Down-sampling, the characteristic pattern after dimensionality reduction is input to third layer.384 third layer, the 4th layer of use convolution kernels, convolution kernel window are big
Small is 3 × 3 pixels, and convolution kernel window carries out convolution operation with 1 to be spaced in the output of respective preceding layer, exports 384
Characteristic pattern, each characteristic pattern connection ReLU non-linear layers do nonlinear transformation, the characteristic pattern after conversion are input into next layer.The
Five layers of use, 256 convolution kernels, convolution kernel window size are 3 × 3 pixels, and convolution kernel window is that interval inputting with 1
Image scrolling product operation, exports 256 characteristic patterns, and each characteristic pattern connection ReLU non-linear layers do nonlinear transformation, will converted
Characteristic pattern afterwards is input to next layer.Five layers of convolutional neural networks of the above are RPN and Fast-RCNN shared convolutional layer.
RPN network struction units:For building RPN convolutional neural networks;RPN convolutional neural networks by six convolutional layers,
Two full articulamentums and a Softmax layer are formed.After building five layers of convolutional neural networks, RPN convolutional neural networks need structure
Layer 6 convolutional layer is built as feature extraction layer.Layer 6 structure is consistent with layer 5, exports 256 characteristic patterns, and characteristic pattern is done
Layer 7 and the 8th layer are input to after nonlinear transformation.Reference frame is built, for each position of infrared image, it is contemplated that 9
Individual possible window:Three kinds of areas { 1282,2562,5122 }, three kinds of ratios { 1:1,1:2,2:1}.Layer 7 and the 8th layer are
Full articulamentum at the same level, layer 7 are the recurrence layer of 24 neurons, the 8th layer be 48 neurons classification layer, after layer 7
Face reuses a Softmax layer.In six convolutional layers, first five layer is characterized extract layer, and layer 6 is characterized mapping layer.7th
The responsible classification of layer, obtains the object candidate area in infrared image;8th layer of responsible recurrence, export the confidence of object candidate area
Degree.Adjustment is normalized to the confidence level of the 8th layer of output in Softmax layers.It is finally defeated according to the confidence level of object candidate area
Go out the location parameter of reliable object candidate area.
Fast-RCNN network struction units:For building Fast-RCNN convolutional neural networks;Fast-RCNN convolutional Neurals
Network is made up of five convolutional layers, a ROI ponds layer, several full articulamentums and a Softmax layer;According to Faster-
RCNN algorithmic characteristics, Fast-RCNN five convolutional layers as and RPN five feature extraction layers be shared convolutional layer.Structure
After five layers of convolutional neural networks, Fast-RCNN uses a ROI ponds layer, ROI layers with 256 characteristic patterns of layer 5 and
The object candidate area of RPN convolutional neural networks output is input, and object candidate area is mapped on characteristic pattern, forms 6 × 6
The characteristic pattern output of individual pixel.It is that several full articulamentums do nonlinear transformation to feature after the layer of ROI ponds, reuses two afterwards
Full articulamentum at the same level, the neuronal quantity of full articulamentum is 4096.With RPN convolutional neural networks, classification layer output target is waited
The location parameter of favored area, nonlinear transformation is done to feature using several full articulamentums, after nonlinear transformation, there is the complete of peer
Articulamentum:Classify and layer and return layer, classification layer is the full connection of 8 neurons, returns the full connection that layer is 2 neurons, ROI
The confidence level of layer output object candidate area is returned after the layer of pond.Return layer and reuse a Softmax layer, Softmax layers are right
Adjustment is normalized in confidence level, and the location parameter of combining target candidate region obtains object candidate area, as whole Fast-
The output of RCNN convolutional neural networks.
Model training module:For being instructed respectively to RPN convolutional neural networks and Fast-RCNN convolutional neural networks
Practice;Specifically, model training module includes RPN training units, Fast-RCNN training units and iterative calculation unit.
RPN training units:For training RPN convolutional neural networks;Specifically training method is:
1st, with the training parameter in different small random number (0~1) initialization RPN convolutional neural networks;
2nd, training sample is input in RPN convolutional neural networks, carries out propagated forward, computational discrimination classification and rectangle frame
The Softmax penalty values of label, the SmoothL1 penalty values of rectangle frame where calculating object candidate area parameter and equipment;
3rd, using stochastic gradient descent algorithm, each layer of cost function Grad in RPN convolutional neural networks is calculated respectively,
The training parameter in RPN convolutional neural networks is updated with Grad;
4th, using backpropagation BP algorithm, each layer in RPN convolutional neural networks of error-sensitivity is calculated, with error spirit
Sensitivity adjusts the training parameter of RPN convolutional neural networks, makes Softmax penalty values and SmoothL1 penalty values minimum, thus
To the RPN convolutional neural networks trained;
5th, the RPN convolutional neural networks for training training sample input, obtain initial target candidate region.
Fast-RCNN training units:For training Fast-RCNN convolutional neural networks, specific training and RPN convolutional Neurals
The training method of network is identical, and it inputs the initial target candidate region and sample data for the output of RPN convolutional neural networks, by
This Fast-RCNN convolutional neural networks trained, and obtain the network parameter of five layers of shared convolutional layer;
Iterate to calculate unit:For reinitializing RPN convolutional Neural nets using the network parameter of five layers of shared convolutional layer
Network, training sample re -training RPN convolutional neural networks are inputted, and export new object candidate area;Meanwhile iterate to calculate single
Member also reinitializes Fast-RCNN convolutional neural networks using the network parameter of five layers of shared convolutional layer, and uses new mesh
Candidate region and sample data are marked as input, re -training Fast-RCNN convolutional neural networks, the Fast- trained
Each layer network parameter of RCNN convolutional neural networks.
Model measurement module:For infrared image to be measured to be inputted into convolutional neural networks model, the result of test model is accurate
True rate.
Target identification module:For the convolutional neural networks for training and pass through test infrared image input to be detected
Model, the power equipment and temperature scale gone out by convolutional neural networks Model Identification in infrared image, and and mark out respectively
Power equipment target area and temperature scale target area.
Temperature detecting module:For the power equipment and temperature scale for recognizing, carried out using image processing techniques
Compare, obtain the temperature results of power equipment.Specifically, the mode of temperature detecting module detection temperature result includes:
1st, environment temperature during collection infrared image is determined, and determines the temperature scale in infrared image;
2nd, temperature maximum region and temperature minimum value region are distinguished according to the position of temperature scale target area, led to
Cross optical recognition program tesseract and read numeral in temperature maximum region and temperature minimum value region respectively, obtaining should
Temperature maximum Tmax and temperature minimum value Tmin in infrared image;
3rd, the value matrix of temperature scale target area tri- values of R, G, B is calculated, it is same in the centre position phase selection of three matrixes
Row, form a rgb value matrix, and the bound rgb value of temperature scale is all identical in every width infrared image;
4th, because temperature scale target area can not be accurately positioned the bound of temperature scale, according to the gradient of R, G, B value
The border rgb value of change and temperature scale determines the bound of temperature scale in each row of R, G, B, removes the upper of rgb value matrix
Lower redundance, obtain new rgb value matrix;
5th, rgb value and temperature are generated with reference to table according to new rgb value matrix and temperature maximum and temperature minimum value;
6th, maximum temperature of the rgb value maximum pixel point as power equipment in power equipment callout box is extracted, by electric power
The maximum temperature and rgb value of equipment are compared with temperature with reference to the rgb value in table, obtain the temperature results of power equipment;
In power equipment target area, because the heating temp of power equipment can be higher than environment temperature, the power equipment that hot stall be present
More obvious, therefore, the rgb value peak that can extract in power equipment target area is read as the temperature of the power equipment
Point, contrasted using the rgb value and rgb value of the point with temperature with reference to the rgb value in table, you can obtain the temperature knot of the power equipment
Fruit.
Fault diagnosis module:For according to device type, using《DL/T 664-2008 charging equipment infrared diagnostics applications
Specification》In standard the temperature results of power equipment are diagnosed, judge whether power equipment hot stall occurs.According to temperature
Detection module obtains the maximum temperature T of power equipment, and normal temperature Tn, Te according to corresponding to device type obtains power equipment are
Environment temperature during infrared image is gathered, is provided in infrared image.Thus the temperature rise value Tr=T-Te of power equipment is obtained, electricity
The temperature approach Td=T-Tn of power equipment, relative temperature difference value δ=(T-Tn)/(T-Te) of power equipment.Temperature rise to power equipment
Value, temperature approach and relative temperature difference value, use《DL/T 664-2008 charging equipment infrared diagnostics application specifications》In standard carry out
Diagnosis, from judging whether the power equipment hot stall occurs.
Fig. 7 is the hardware device structural representation of the power equipment thermal fault detection method of the embodiment of the present invention, such as Fig. 7 institutes
Show, the equipment includes one or more processors and memory.By taking a processor as an example, the equipment can also include:It is defeated
Enter device and output device.
Processor, memory, input unit and output device can be connected by bus or other modes, in Fig. 7 with
Exemplified by being connected by bus.
Memory as a kind of non-transient computer readable storage medium storing program for executing, available for store non-transient software program, it is non-temporarily
State computer executable program and module.Processor is by running storage non-transient software program in memory, instruction
And module, so as to perform the various function application of electronic equipment and data processing, that is, realize the place of above method embodiment
Reason method.
Memory can include storing program area and storage data field, wherein, storing program area can storage program area, extremely
Application program required for few One function;Storage data field can data storage etc..In addition, memory can be included at a high speed at random
Memory is accessed, can also include non-transient memory, a for example, at least disk memory, flush memory device or other are non-
Transient state solid-state memory.In certain embodiments, memory is optional including relative to the remotely located memory of processor, this
A little remote memories can pass through network connection to processing unit.The example of above-mentioned network includes but is not limited to internet, enterprise
In-house network, LAN, mobile radio communication and combinations thereof.
Input unit can receive the numeral or character information of input, and produce signal input.Output device may include to show
The display devices such as display screen.
One or more of modules are stored in the memory, when by one or more of computing devices
When, perform the following operation of any of the above-described embodiment of the method:
The infrared image of power equipment is gathered, convolutional neural networks model is built according to the infrared image;
Infrared image to be detected is inputted into the convolutional neural networks model, known by the convolutional neural networks model
The temperature scale and power equipment not gone out in infrared image;
According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature reference
Table, and the rgb value of identified power equipment is extracted, by the rgb value of extraction and the rgb value with temperature with reference to the RGB in table
Value is compared, and obtains the temperature results of identified power equipment;
Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judge that the electric power is set
It is standby whether hot stall occur.
The said goods can perform the method that the embodiment of the present invention is provided, and possesses the corresponding functional module of execution method and has
Beneficial effect.Not ins and outs of detailed description in the present embodiment, reference can be made to method provided in an embodiment of the present invention.
The embodiments of the invention provide a kind of non-transient (non-volatile) computer-readable storage medium, the computer storage is situated between
Matter is stored with computer executable instructions, the executable following operation of the computer executable instructions:
The infrared image of power equipment is gathered, convolutional neural networks model is built according to the infrared image;
Infrared image to be detected is inputted into the convolutional neural networks model, known by the convolutional neural networks model
The temperature scale and power equipment not gone out in infrared image;
According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature reference
Table, and the rgb value of identified power equipment is extracted, by the rgb value of extraction and the rgb value with temperature with reference to the RGB in table
Value is compared, and obtains the temperature results of identified power equipment;
Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judge that the electric power is set
It is standby whether hot stall occur.
The embodiments of the invention provide a kind of computer program product, the computer program product is non-temporary including being stored in
Computer program on state computer-readable recording medium, the computer program include programmed instruction, when described program instructs
When being computer-executed, the computer is set to perform following operate:
The infrared image of power equipment is gathered, convolutional neural networks model is built according to the infrared image;
Infrared image to be detected is inputted into the convolutional neural networks model, known by the convolutional neural networks model
The temperature scale and power equipment not gone out in infrared image;
According to the rgb value of the temperature scale pixel identified and temperature scale bound generation rgb value and temperature reference
Table, and the rgb value of identified power equipment is extracted, by the rgb value of extraction and the rgb value with temperature with reference to the RGB in table
Value is compared, and obtains the temperature results of identified power equipment;
Diagnosed according to the temperature results of power equipment of the network system diagnostic criteria to being identified, judge that the electric power is set
It is standby whether hot stall occur.
Power equipment thermal fault detection method, system and the electronic equipment of the embodiment of the present invention are calculated using Faster-RCNN
Method trains a special target identification convolutional neural networks model, and the model can identify figure by the infrared image of input
Power equipment and temperature scale as in, and the temperature of rgb value Comparison Method reading power equipment is used, eventually through《DL/
T664-2008 charging equipment infrared diagnostics application specifications》In standard judge whether power equipment hot stall occurs.The present invention is logical
Cross convolution neural network model efficiently, accurately identify power equipment, temperature is accurately read by rgb value, it is final efficiently, it is reliable
Diagnosing electric power hot stall.The present invention has reformed the mode of power equipment thermal fault detection, lifts the intelligence of network system
Energyization is horizontal.
The foregoing description of the disclosed embodiments, professional and technical personnel in the field are enable to realize or using the present invention.
A variety of modifications to these embodiments will be apparent for those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, it is of the invention
The embodiments shown herein is not intended to be limited to, and is to fit to and principles disclosed herein and features of novelty phase one
The most wide scope caused.