CN113486877A - Power equipment infrared image real-time detection and diagnosis method based on lightweight artificial intelligence model - Google Patents
Power equipment infrared image real-time detection and diagnosis method based on lightweight artificial intelligence model Download PDFInfo
- Publication number
- CN113486877A CN113486877A CN202110640460.0A CN202110640460A CN113486877A CN 113486877 A CN113486877 A CN 113486877A CN 202110640460 A CN202110640460 A CN 202110640460A CN 113486877 A CN113486877 A CN 113486877A
- Authority
- CN
- China
- Prior art keywords
- model
- infrared image
- electrical equipment
- real
- data set
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- 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
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Radiation Pyrometers (AREA)
Abstract
The invention provides a power equipment infrared image real-time detection and diagnosis method based on a lightweight artificial intelligence model, which comprises the following steps: forming an effective data set by collecting and processing normal and fault electrical equipment infrared images; constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model; training the model on a deep learning platform, and storing the trained model; detecting and diagnosing a target on the acquired infrared image by using an improved artificial intelligence model; performing fault diagnosis on the image diagnosed as the fault by using an electrical equipment fault diagnosis method combined with a K mean value clustering algorithm, detecting a temperature highest point and extracting the temperature of the point; the infrared image of various electrical equipment of the transformer substation can be automatically detected and diagnosed. The invention can greatly reduce the calculation amount, thereby being capable of being deployed in embedded equipment, realizing the effective detection of the electrical equipment and simultaneously meeting the requirement of real-time detection.
Description
Technical Field
The invention relates to the field of safety monitoring of the running state of electric equipment, in particular to an infrared image real-time detection and diagnosis method for electric equipment based on a lightweight artificial intelligence model.
Background
With the rapid development of power systems, people have higher dependence on electric power and larger demand for electric power. Today, power systems face significant challenges to meet the ever-increasing demand for sustainable energy, and therefore larger, more complex power systems are needed, but this complexity affects the task of inspecting and maintaining them. The transformer substation is used as an important part in the operation of a modern power system and a future intelligent power grid, plays a key role in the safe operation of the whole power grid, and is also key in accident prediction on the system inspection of the transformer substation. Because when an electrical problem occurs in the substation, unnecessary energy loss may be caused, and a costly system shutdown may be caused, even a technician is injured. Therefore, in order to minimize the failure of the power equipment and avoid a great economic loss due to power outage, continuous inspection and preventive maintenance are required, and early detection of an initial failure is required to prevent the occurrence of a permanent failure, thereby ensuring safe long-term operation.
Infrared thermography has become a widely accepted condition monitoring technique because it has many advantages over other types of sensors. The infrared thermal image detection is a technology for diagnosing whether the running state of the equipment is good or not based on the thermal distribution state of the equipment, and the technology can be in non-contact distance with the detected equipment and has wide temperature measurement range so as to rapidly carry out scanning detection. Since infrared detection of electrical devices generates a large amount of picture data, analyzing infrared images manually to detect the status of electrical devices may consume a lot of time and effort, and may also result in erroneous diagnosis results. And the current classic machine learning algorithm is difficult to effectively identify the infrared image fault abnormal heating point of the power transmission and transformation equipment.
In recent years, with the improvement of computer computing power, deep learning has received attention from more and more researchers. The deep learning method is more and more widely applied to the aspects of image classification, fault diagnosis, target detection and the like. For example: wang, m.dong, m.ren, z.y.wu, c.x.guo, t.x.zhuang, o.pischler and j.c.xie.automatic fault diagnosis of infracted insulator images based on image analysis and Measurement, vol.69, No.8, pp.5345-5355, and aug.2020, which proposes an automatic diagnostic method for infrared insulator image example segmentation and temperature analysis based on mass R-CNN. The document provides an automatic positioning, identifying and diagnosing method of external electric insulating equipment based on YOLOv3, and the method is used for extracting image data characteristics under an insulator visible light channel. Liuyunpeng, Figliotong, Wujianhua, Jixin and Lihui, the infrared image target detection method for the abnormal heating point of the power transmission and transformation equipment based on deep learning [ J ] southern power grid technology, Vol.13, No.2, pp.27-33, Feb.2019, the document provides a method for realizing detection, identification and positioning of the infrared image heating fault of the power transmission and transformation based on the fast RCNN algorithm.
The corresponding patent also comprises a CN202010905683.0 method for identifying the transformer substation insulator infrared image detection model in any direction based on artificial intelligence and a CN202010906630.0 method for detecting the transformer substation insulator infrared image based on artificial intelligence.
Although the method has a good effect on detection precision, better balance research is not carried out on the size of a model, the detection speed and the detection precision, and effective judgment on the running state of the electrical equipment is not realized in a limited environment, so that the method can be deployed in the limited environment after balance research.
Disclosure of Invention
The invention aims to solve the technical problem of providing a power equipment infrared image real-time detection and diagnosis method based on a lightweight artificial intelligence model, which reduces the size of the model through model lightweight, so that an effective electrical equipment detection and diagnosis model can be deployed in a limited environment (such as an unmanned aerial vehicle detector or embedded equipment such as a handheld camera detector), effective detection on various electrical equipment can be realized, the requirement of real-time detection is met, and operation resources are effectively utilized. The method has universality and effectiveness, and ensures safe and real-time automatic detection and diagnosis of the electrical equipment of the transformer substation.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows: a real-time detection and diagnosis method for infrared images of electric equipment based on a lightweight artificial intelligence model comprises the following steps:
s1, forming an effective data set by collecting and processing infrared images of normal and fault electrical equipment;
constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
training the model on a deep learning platform, and storing the trained model;
s2, detecting and diagnosing the target of the collected infrared image by using the improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
s3, carrying out fault diagnosis on the image diagnosed as the fault in the S2 by using an electrical equipment fault diagnosis method combined with a K-means clustering algorithm, detecting a temperature highest point and extracting the temperature of the point;
the infrared images of various electrical equipment of the transformer substation are automatically detected and diagnosed through the steps.
In a preferred embodiment, the step S1 is divided into the following steps:
s01, acquiring an infrared image of the electrical equipment of the transformer substation;
s02, preprocessing the acquired image through an algorithm to form a data set for training;
s03, performing target label processing on the acquired normal electrical equipment data set and the acquired fault electrical equipment data set;
s04, randomly distributing the processed data set into a training set and a testing set;
training the model together for 200000 steps, wherein the data input size is 300 × 300 pixels, 16 pictures are trained in one batch, the learning rate is set to be 0.001, the momentum is set to be 0.9, random gradient descent is used as an optimization algorithm, and the weight attenuation is 0.0005;
s05, initially constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
s06, adjusting and training the parameters of the model by using the divided training set;
s07, detecting and diagnosing the target of the trained detection and diagnosis model by using the divided test set;
and S08, carrying out fault diagnosis on the pictures diagnosed as the faults by combining a K mean value clustering algorithm, detecting the highest temperature point and extracting the temperature of the point so as to verify the effectiveness of the model.
In a preferred embodiment, in step S03:
performing data expansion on the original infrared image by adopting luminosity distortion methods such as random brightness, contrast, hue, saturation and random noise adjustment and geometric distortion methods such as random turning, translation, scaling and rotation on the acquired data set, or taking the three-dimensional forms of various electrical equipment as models to obtain training picture sets in multiple directions and angles to form a data set applied to the models;
marking various electrical devices in the data set through a frame selection operation;
marking the equipment by a normal electrical equipment data set through software or algorithm marks, and marking a heating fault point by a fault electrical equipment data set;
in a preferred scheme, the training is performed by taking the shape of the electrical equipment as a recognition feature, and the training method comprises the following steps of:
s11, taking the direction of the electrical equipment as the direction vector of the picture of the infrared image data set of the electrical equipment, and preprocessing the picture in one or more image processing modes to expand the data set so as to keep the direction of the electrical equipment in the picture consistent;
and S12, adopting a VGG16 structure, identifying the shapes of single or multiple electrical equipment as label areas, expanding the label areas in a proportional mode, then using the image areas where the electrical equipment is located as label areas, carrying out intelligent identification and framing to manufacture a data set.
In the preferred scheme, the backbone network structure of the light-weight FSSD is an improved light-weight model SqueezeNet structure;
s21, deleting Conv10 and a global maximum pooling layer on the basis of the Squeezenet network;
s22, replacing the VGG16 structure by the modified structure to serve as a backbone network of the improved light-weight FSSD;
meanwhile, in order to compensate for the influence of light weight on detection precision, a plurality of convolution layers with gradually reduced sizes are added behind a backbone network, and then a plurality of bypass connections are added in the backbone network.
In a preferred embodiment, in step S21: the structure of complex connecting branches combining 1 × 1 convolution by applying residual connection is adopted to enhance the propagation of features and a plurality of convolution layers are added behind a backbone network, so that the influence of model lightweight on detection accuracy is reduced.
In a preferred scheme, the light-weight FSSD adopts a feature enhancement module to improve the capability of extracting the infrared image features of the electrical equipment. The feature enhancement module adopts a multi-branch feature extraction and aggregation mode to increase semantic information. Each branch is firstly subjected to dimensionality reduction through a 1 x 1 convolution kernel, then convolution kernels with different sizes are adopted to extract features, 5 x 5 convolution kernels in the branches are replaced by two 3 x 3 convolution cascade modes for reducing calculation amount, meanwhile, space separation convolution is adopted, two channels are set to be one of the channels, one of the channels is subjected to 1 x 3 convolution firstly, then the other channel is subjected to 3 x 1 convolution, and then the other channel is subjected to 1 x 3 convolution. The above operation reduces the amount of calculation of the model while increasing the diversity of feature extraction. And the feature enhancement module uses batch normalization operations on top of each convolution layer to speed up the convergence of the model. A residual error connecting branch circuit combined with an attention mechanism is added into the module to improve the detection effect, so that the characteristic information is more effectively extracted. And finally, using the hole convolution with different expansion coefficients in the last convolution kernel of each branch, wherein the purpose of the structure is to enable the network to capture more characteristic information, widen the receptive field of the characteristics and simultaneously keep the same number of parameters.
In the preferred scheme, the problem of neuron death can be solved only by adding a very small amount of parameters to the PReLU, so that the PReLU is used for replacing the ReLU as an activation function, the problem of gradient explosion is avoided by applying batch normalization, the convergence of the model is improved, and the recognition rate and the detection speed of the model are improved; the label aspect ratio in the infrared image data set of the electrical equipment is clustered by using a k-means + + clustering algorithm, so that a blind search mechanism of a default frame is replaced, and the self-adaptive change of the model default frame is realized. By using the clustering method, a better clustering center can be selected, so that the problem of clustering initialization is solved, and clustering errors are reduced.
In a preferred scheme, the model obtains prior frames with different scales in different level feature maps, calculates the position loss and the confidence loss of a default frame obtained by matching, and adopts a weighted sum form of the confidence loss and the position loss as a total target loss function, wherein the model loss function is as follows:
wherein, x takes 0 or 1 to represent whether the prior frame is matched with the real label frame, c represents the category confidence, l represents the real information of the prediction frame, g represents the real information of the real label frame, N represents the number of matched default frames, and alpha represents the weight of the two frames; the confidence penalty is SoftMax loss and the location penalty is the smooth-L1 penalty between the prior box and the real tag box parameters.
In the preferred scheme, the highest temperature point is detected, the temperature of the point is extracted, and the fault degree of the target equipment is determined by combining the temperature matrix of the equipment and the environmental temperature information.
Power equipment infrared image real-time detection and diagnosis method based on lightweight artificial intelligence model
According to the infrared image real-time detection and diagnosis method for the power equipment based on the lightweight artificial intelligence model, the lightweight artificial intelligence model is adopted for processing, detecting and diagnosing, the calculation amount is greatly reduced, so that an effective electric equipment detection and diagnosis model can be deployed in a limited environment such as an embedded type equipment, effective detection on the electric equipment can be realized, the requirement of real-time detection is met, the detection and identification efficiency is improved, the calculation resources are effectively utilized, and the safety and real-time automatic detection and diagnosis of the electric equipment of a transformer substation are ensured. By improving the structure of the SqueezeNet, a structure of a complex connecting branch is adopted to enhance the propagation of characteristics and a plurality of convolution layers are added behind a main network, so that the influence of model lightweight on detection precision is reduced. And the application optimized feature enhancement module is used for improving the capability of extracting the infrared image features of the electrical equipment. Further optimization of these functions by using a channel attention mechanism helps to extract feature information more efficiently. Meanwhile, the activation function of the model is improved, and the problem of gradient explosion is avoided by applying batch normalization, so that the convergence of the model is better, and the recognition rate and the detection speed of the model are improved. Furthermore, the aspect ratio information of the targets in the data set is counted by using a cluster analysis method, and the self-adaptive change of the default frame is realized, so that the detection performance of the model on the interested targets is enhanced. Training a training set input model by adopting a strategy of model weight random initialization, and adjusting parameters according to a training result through detailed experiments so as to confirm an optimal lightweight model; in a further preferred scheme, the diversity of data can be improved by carrying out different pre-processing on the data sets acquired on site, and the over-fitting training is prevented. The invention has universality and effectiveness.
Drawings
The invention is further illustrated by the following examples in conjunction with the accompanying drawings:
FIG. 1 is a schematic flow chart of the present invention.
Fig. 2 is a data set image of a portion of a normal electrical device of the present invention.
FIG. 3 is a data set image of a partially failed electrical device of an embodiment of the present invention.
Fig. 4 is a structural diagram of an improved light-weight FSSD power equipment infrared image real-time detection and diagnosis model according to an embodiment of the present invention.
Fig. 5 is a diagram of a feature enhancement module according to an embodiment of the present invention.
Fig. 6 is an aspect ratio clustering result diagram of various electrical devices in a data set label according to an embodiment of the present invention.
Fig. 7 is a diagram illustrating an effect of detecting an infrared image of a part of the test current collector in the embodiment of the present invention.
FIG. 8 is a graph illustrating the testing effect of a partially failed electrical device data set according to an embodiment of the present invention.
Detailed Description
Example 1:
as shown in fig. 1, a method for real-time detection and diagnosis of infrared images of electrical equipment based on a lightweight artificial intelligence model is characterized by comprising the following steps:
the method for detecting and diagnosing the infrared images of the electrical equipment in real time based on the lightweight artificial intelligence model S1 is characterized in that an effective data set is formed by collecting and processing normal and fault electrical equipment infrared images;
constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
training the model on a deep learning platform, and storing the trained model;
s2, detecting and diagnosing the target of the collected infrared image by using the improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
and S3, carrying out fault diagnosis on the image diagnosed as the fault in the S2 by using an electrical equipment fault diagnosis method combined with a K-means clustering algorithm, detecting the highest temperature point and extracting the temperature of the point.
The infrared images of various electrical equipment of the transformer substation are automatically detected and diagnosed through the steps.
In a preferred embodiment, the step S1 is divided into the following steps:
s01, acquiring an infrared image of the electrical equipment of the transformer substation; as shown in fig. 2 and 3. The acquired infrared images of the various electrical equipment are obtained by a substation technician by taking a picture on site through a handheld thermal infrared imager or by a patrol robot carrying the thermal infrared imager in the substation; the five kinds of electrical equipment are respectively a lightning arrester, a circuit breaker, an isolating switch, a mutual inductor and an insulator.
S02, preprocessing the acquired image through an algorithm to form a data set for training; in an optional scheme, the acquired data set, including a normal electrical device data set and a fault electrical device data set, is subjected to data expansion on the original infrared image by using photometric distortion methods such as random brightness, contrast, hue, saturation, random noise adjustment and geometric distortion methods such as random inversion, translation, scaling and rotation to form a data set applied to the model.
Marking various electrical devices in the data set through a frame selection operation;
marking the equipment by a normal electrical equipment data set through software or algorithm marks, and marking a heating fault point by a fault electrical equipment data set; and finally, manufacturing a data set to be trained and detected.
In another alternative scheme, the 3d forms of various electrical devices are used as models to obtain training atlas with multiple directions and angles, and the shapes of various electrical devices are used as recognition features to train, and the method comprises the following steps:
s11, taking the direction of the electrical equipment as the direction vector of the picture of the infrared image data set of the electrical equipment, preprocessing the picture in one or more image processing modes to expand the data set so as to keep the direction of the electrical equipment in the picture approximately consistent;
s12, adopting a VGG16 structure, taking the shapes of single or multiple electrical devices as a label area for recognition, specifically, making a 3-dimensional model according to the shapes of the electrical devices, projecting the single or multiple models in different directions to obtain the model as a training set, thereby extracting the shape features of the electrical devices and facilitating the rapid recognition of the electrical devices in a complex background. For example, each projected feature of a circular truncated cone, and projected features of a plurality of consecutive suspected patterns, and projected features at locations of increasing diameter in a linear pattern. And expanding the recognized images in a proportional mode, then taking the image areas where various electrical equipment are located as label areas, carrying out intelligent recognition and framing to manufacture a data set to be trained and detected. By intelligently identifying the made label area, the consumption of operation resources is greatly reduced, and the efficiency is improved. The electric equipment has obvious shape characteristics and is easy to identify from the image, even if noise exists after identification, the heat generation of the noise is almost negligible, so that the occupation of computing resources for final detection is small and can be ignored. Through the processing of the step, the detection efficiency is further improved.
S03, performing target label processing on the acquired normal electrical equipment data set and the acquired fault electrical equipment data set;
s04, randomly distributing the processed data set into a training set and a testing set; preferably, the number of training sets is greater than the number of test sets. Preferably, 80% of the data set is divided into the training set and 20% into the test set.
Due to the adoption of a model weight random initialization strategy, the model is trained for 200000 steps, the data input size is 300 multiplied by 300 pixels, 16 pictures are trained in one batch, the learning rate is set to be 0.001, the momentum is set to be 0.9, random gradient descent is used as an optimization algorithm, and the weight attenuation is 0.0005;
s05, initially constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
s06, adjusting and training the parameters of the model by using the divided training set;
s07, detecting and diagnosing the target of the trained detection and diagnosis model by using the divided test set;
and S08, carrying out fault diagnosis on the pictures diagnosed as the faults by combining a K mean value clustering algorithm, detecting the highest temperature point and extracting the temperature of the point so as to verify the effectiveness of the model.
In a preferred embodiment, in step S03: as shown in fig. 4, the backbone network structure of the lightweight FSSD is an improved lightweight model SqueezeNet structure;
s21, deleting Conv10 and a global maximum pooling layer on the basis of the Squeezenet network;
s22, replacing the VGG16 structure by the modified structure, and using the modified structure as an improved main network of the single-shot multi-box detector to lighten the main network of the FSSD;
in order to compensate the influence of light weight on the detection precision, a preferable scheme is that an improved light weight model SqueezeNet structure is adopted in a trunk part of an improved model, a complex connecting branch structure is adopted to enhance the characteristic propagation, and a plurality of convolution layers are added behind a trunk network, so that the influence of the light weight of the model on the detection precision is reduced.
In another alternative, in step S21: the structure of complex connecting branches combining 1 × 1 convolution by applying residual connection is adopted to enhance the propagation of features and a plurality of convolution layers are added behind a backbone network, so that the influence of model lightweight on detection accuracy is reduced.
In a preferred scheme, the light-weight FSSD adopts a feature enhancement module to improve the capability of extracting the infrared image features of the electrical equipment. The feature enhancement module adopts a multi-branch feature extraction and aggregation mode to increase semantic information. Each branch is firstly subjected to dimensionality reduction through a 1 x 1 convolution kernel, then convolution kernels with different sizes are adopted to extract features, 5 x 5 convolution kernels in the branches are replaced by two 3 x 3 convolution cascade modes for reducing calculation amount, meanwhile, space separation convolution is adopted, two channels are set to be one of the channels, one of the channels is subjected to 1 x 3 convolution firstly, then the other channel is subjected to 3 x 1 convolution, and then the other channel is subjected to 1 x 3 convolution. The above operation reduces the amount of calculation of the model while increasing the diversity of feature extraction. And the feature enhancement module uses batch normalization operations on top of each convolution layer to speed up the convergence of the model. A residual error connecting branch circuit combined with an attention mechanism is added into the module to improve the detection effect, so that the characteristic information is more effectively extracted. And finally, using the hole convolution with different expansion coefficients in the last convolution kernel of each branch, wherein the purpose of the structure is to enable the network to capture more characteristic information, widen the receptive field of the characteristics and simultaneously keep the same number of parameters.
In the preferred scheme, the problem of neuron death can be solved only by adding a very small amount of parameters to the PReLU, so that the PReLU is used for replacing the ReLU as an activation function in the embodiment, and the problem of gradient explosion is avoided by applying batch normalization, so that the convergence of the model is improved, and the recognition rate and the detection speed of the model are improved; in the embodiment, a k-means + + clustering algorithm is used for clustering the aspect ratio of the labels in the infrared image data set of the electrical equipment, so that a blind search mechanism of a default frame is replaced, and the self-adaptive change of the model default frame is realized. By using the clustering method, a better clustering center can be selected, so that the problem of clustering initialization is solved, and clustering errors are reduced.
In the preferred scheme, the model obtains prior frames with different scales from feature maps at different levels, and calculates the position loss and the confidence loss of a default frame obtained by matching. The total target loss function takes the form of a weighted sum of confidence loss and position loss, and the model loss function is as follows:
wherein, x takes 0 or 1 to represent whether the prior frame is matched with the real label frame, c represents the category confidence, l represents the real information of the prediction frame, g represents the real information of the real label frame, N represents the number of matched default frames, and alpha represents the weight of the two frames; the confidence penalty is SoftMax loss and the location penalty is the smooth-L1 penalty between the prior box and the real tag box parameters.
In the preferred scheme, the highest temperature point is detected, the temperature of the point is extracted, and the fault degree of the target equipment is determined by combining the temperature matrix of the equipment and the environmental temperature information.
The results of the model test are shown in fig. 7, and the detection has good effect. The whole test set is tested, the final recognition average accuracy of the five electrical devices is respectively 95.26% of the lightning arrester, 90.76% of the circuit breaker, 88.95% of the isolating switch, 90.36% of the mutual inductor, 87.45% of the insulator and 90.56% of the average accuracy average value of the whole test set, the size of the model is 97.1MB, and the time for detecting a single picture is 28.470 ms. The example shows that the method can accurately identify various electrical equipment while realizing the lightweight model, and provides a real-time and reliable basis for judging the detection working state of the electrical equipment.
The above-described embodiments are merely preferred embodiments of the present invention, and should not be construed as limiting the present invention, and features in the embodiments and examples in the present application may be arbitrarily combined with each other without conflict. The protection scope of the present invention is defined by the claims, and includes equivalents of technical features of the claims. I.e., equivalent alterations and modifications within the scope hereof, are also intended to be within the scope of the invention.
Claims (10)
1. A real-time detection and diagnosis method for infrared images of electrical equipment based on a lightweight artificial intelligence model is characterized by comprising the following steps:
s1, forming an effective data set by collecting and processing infrared images of normal and fault electrical equipment;
constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
training the model on a deep learning platform, and storing the trained model;
s2, detecting and diagnosing the target of the collected infrared image by using the improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
s3, carrying out fault diagnosis on the image diagnosed as the fault in the S2 by using an electrical equipment fault diagnosis method combined with a K-means clustering algorithm, detecting a temperature highest point and extracting the temperature of the point;
the infrared images of various electrical equipment of the transformer substation are automatically detected and diagnosed through the steps.
2. The method for detecting and diagnosing the infrared image of the electric power equipment in real time based on the lightweight artificial intelligence model as claimed in claim 1, wherein the step S1 is divided into the following steps:
s01, acquiring an infrared image of the electrical equipment of the transformer substation;
s02, preprocessing the acquired image through an algorithm to form a data set for training;
s03, performing target label processing on the acquired normal electrical equipment data set and the acquired fault electrical equipment data set;
s04, randomly distributing the processed data set into a training set and a testing set;
training the model together for 200000 steps, wherein the data input size is 300 × 300 pixels, 16 pictures are trained in one batch, the learning rate is set to be 0.001, the momentum is set to be 0.9, random gradient descent is used as an optimization algorithm, and the weight attenuation is 0.0005;
s05, initially constructing an improved lightweight FSSD power equipment infrared image real-time detection and diagnosis model;
s06, adjusting and training the parameters of the model by using the divided training set;
s07, detecting and diagnosing the target of the trained detection and diagnosis model by using the divided test set;
and S08, carrying out fault diagnosis on the pictures diagnosed as the faults by combining a K mean value clustering algorithm, detecting the highest temperature point and extracting the temperature of the point.
3. The method for detecting and diagnosing the infrared image of the electric power equipment in real time based on the lightweight artificial intelligence model as claimed in claim 2, wherein in step S03:
performing data expansion on the original infrared image by adopting luminosity distortion methods such as random brightness, contrast, hue, saturation and random noise adjustment and geometric distortion methods such as random turning, translation, scaling and rotation on the acquired data set, or taking the three-dimensional forms of various electrical equipment as models to obtain training picture sets in multiple directions and angles to form a data set applied to the models;
marking various electrical devices in the data set through a frame selection operation;
and the normal electrical equipment data set marks the equipment through software or algorithm marks, and the fault electrical equipment data set marks heating fault points.
4. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 3, wherein the method comprises the following steps: the method takes the shape of the electrical equipment as a recognition feature to train, and comprises the following steps:
s11, taking the direction of the electrical equipment as the direction vector of the picture of the infrared image data set of the electrical equipment, and preprocessing the picture in one or more image processing modes to expand the data set so as to keep the direction of the electrical equipment in the picture consistent;
and S12, adopting a VGG16 structure, identifying the shapes of single or multiple electrical equipment as label areas, expanding the label areas in a proportional mode, then using the image areas where the electrical equipment is located as label areas, carrying out intelligent identification and framing to manufacture a data set.
5. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 1, wherein the method comprises the following steps: the main network structure of the light weight FSSD is an improved light weight model SqueezeNet structure;
s21, deleting Conv10 and a global maximum pooling layer on the basis of the Squeezenet network;
s22, replacing the VGG16 structure by the modified structure to serve as a backbone network of the improved light-weight FSSD;
meanwhile, in order to compensate for the influence of light weight on detection precision, a plurality of convolution layers with gradually reduced sizes are added behind a backbone network, and then a plurality of bypass connections are added in the backbone network.
6. The method for real-time detection and diagnosis of infrared images of electric power equipment based on the lightweight artificial intelligence model as claimed in claim 5, wherein in step S21: the structure of complex connecting branches combining 1 × 1 convolution by applying residual connection is adopted to enhance the propagation of features and a plurality of convolution layers are added behind a backbone network, so that the influence of model lightweight on detection accuracy is reduced.
7. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 5, wherein the method comprises the following steps: the light FSSD adopts a feature enhancement module to improve the capability of extracting the infrared image features of the electrical equipment; the feature enhancement module adopts a multi-branch feature extraction and aggregation mode to increase semantic information; each branch is firstly subjected to dimension reduction through a 1 × 1 convolution kernel, then convolution kernels with different sizes are adopted to extract features, 5 × 5 convolution kernels in the branches are replaced by two 3 × 3 convolution cascade modes for reducing calculation amount, meanwhile, space separation convolution is adopted, two channels are set to be one of the two channels, and the two channels are firstly subjected to 1 × 3 convolution and then subjected to 3 × 1 convolution; the other channel is convolved 3 x 1 and then 1 x 3.
8. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 5, wherein the method comprises the following steps: the PReLU is used for replacing the ReLU to serve as an activation function, batch normalization is applied to avoid the problem of gradient explosion, the convergence of the model is improved, and the recognition rate and the detection speed of the model are improved;
and clustering the aspect ratio of the labels in the infrared image data set of the electrical equipment by using a k-means + + clustering algorithm so as to replace a blind search mechanism of a default frame and realize the self-adaptive change of the model default frame.
9. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 5, wherein the method comprises the following steps: the model obtains prior frames with different scales on different level feature maps, calculates the position loss and the confidence loss of a default frame obtained through matching, adopts a weighted addition form of the confidence loss and the position loss as a total target loss function, and the model loss function is as follows:
wherein, x takes 0 or 1 to represent whether the prior frame is matched with the real label frame, c represents the category confidence, l represents the real information of the prediction frame, g represents the real information of the real label frame, N represents the number of matched default frames, and alpha represents the weight of the two frames; the confidence penalty is SoftMax loss and the location penalty is the smooth-L1 penalty between the prior box and the real tag box parameters.
10. The infrared image real-time detection and diagnosis method for the power equipment based on the light artificial intelligence model as claimed in claim 1, wherein the method comprises the following steps: and detecting the highest temperature point, extracting the temperature of the point, and determining the fault degree of the target equipment by combining the temperature matrix of the equipment and the environmental temperature information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110640460.0A CN113486877B (en) | 2021-06-08 | 2021-06-08 | Power equipment infrared image real-time detection and diagnosis method based on light artificial intelligent model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110640460.0A CN113486877B (en) | 2021-06-08 | 2021-06-08 | Power equipment infrared image real-time detection and diagnosis method based on light artificial intelligent model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113486877A true CN113486877A (en) | 2021-10-08 |
| CN113486877B CN113486877B (en) | 2023-10-17 |
Family
ID=77934921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110640460.0A Active CN113486877B (en) | 2021-06-08 | 2021-06-08 | Power equipment infrared image real-time detection and diagnosis method based on light artificial intelligent model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113486877B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114034390A (en) * | 2021-11-08 | 2022-02-11 | 山东大学 | Equipment temperature anomaly detection system based on infrared detection |
| CN114689122A (en) * | 2022-03-31 | 2022-07-01 | 国网北京市电力公司 | Equipment fault monitoring method, device, equipment and medium |
| CN117152060A (en) * | 2023-07-25 | 2023-12-01 | 长安大学 | Fractured rock mass fracture quantity determining method based on machine learning, fractured rock mass quality evaluating method, electronic equipment and storage medium |
| 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 |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598736A (en) * | 2019-08-06 | 2019-12-20 | 西安理工大学 | Power equipment infrared image fault positioning, identifying and predicting method |
| CN112200178A (en) * | 2020-09-01 | 2021-01-08 | 广西大学 | Transformer substation insulator infrared image detection method based on artificial intelligence |
| CN112308825A (en) * | 2020-10-22 | 2021-02-02 | 江苏大学 | SqueezeNet-based crop leaf disease identification method |
-
2021
- 2021-06-08 CN CN202110640460.0A patent/CN113486877B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598736A (en) * | 2019-08-06 | 2019-12-20 | 西安理工大学 | Power equipment infrared image fault positioning, identifying and predicting method |
| CN112200178A (en) * | 2020-09-01 | 2021-01-08 | 广西大学 | Transformer substation insulator infrared image detection method based on artificial intelligence |
| CN112308825A (en) * | 2020-10-22 | 2021-02-02 | 江苏大学 | SqueezeNet-based crop leaf disease identification method |
Non-Patent Citations (1)
| Title |
|---|
| 李浩: "基于计算机视觉的电力设备红外检测研究", 《长沙理工大学》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114034390A (en) * | 2021-11-08 | 2022-02-11 | 山东大学 | Equipment temperature anomaly detection system based on infrared detection |
| CN114034390B (en) * | 2021-11-08 | 2023-11-03 | 山东大学 | Equipment temperature anomaly detection system based on infrared detection |
| CN114689122A (en) * | 2022-03-31 | 2022-07-01 | 国网北京市电力公司 | Equipment fault monitoring method, device, equipment and medium |
| CN114689122B (en) * | 2022-03-31 | 2023-11-10 | 国网北京市电力公司 | Equipment fault monitoring method, device, equipment and medium |
| CN117152060A (en) * | 2023-07-25 | 2023-12-01 | 长安大学 | Fractured rock mass fracture quantity determining method based on machine learning, fractured rock mass quality evaluating method, electronic equipment and storage medium |
| CN117152060B (en) * | 2023-07-25 | 2024-06-07 | 长安大学 | Method for determining number of fractured rock mass fractures based on machine learning, electronic equipment and storage medium |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113486877B (en) | 2023-10-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112380952B (en) | Power equipment infrared image real-time detection and identification method based on artificial intelligence | |
| CN113486877B (en) | Power equipment infrared image real-time detection and diagnosis method based on light artificial intelligent model | |
| CN112199993B (en) | Method for identifying transformer substation insulator infrared image detection model in any direction based on artificial intelligence | |
| CN112200178B (en) | Transformer substation insulator infrared image detection method based on artificial intelligence | |
| CN107944396B (en) | Knife switch state identification method based on improved deep learning | |
| CN113159334B (en) | Electrical equipment infrared image real-time detection and diagnosis method based on light-weight deep learning | |
| CN111611874B (en) | Face mask wearing detection method based on ResNet and Canny | |
| CN114240878A (en) | Routing inspection scene-oriented insulator defect detection neural network construction and optimization method | |
| CN112233073A (en) | Real-time detection method for infrared thermal imaging abnormity of power transformation equipment | |
| CN114445706A (en) | Power transmission line target detection and identification method based on feature fusion | |
| CN109376580B (en) | Electric power tower component identification method based on deep learning | |
| CN115690542A (en) | Improved yolov 5-based aerial insulator directional identification method | |
| Wanguo et al. | Typical defect detection technology of transmission line based on deep learning | |
| CN113962973A (en) | Power transmission line unmanned aerial vehicle intelligent inspection system and method based on satellite technology | |
| CN113205039A (en) | Power equipment fault image identification and disaster investigation system and method based on multiple DCNNs | |
| CN116503399A (en) | Insulator pollution flashover detection method based on YOLO-AFPS | |
| CN116580330A (en) | Machine test abnormal behavior detection method based on double-flow network | |
| CN115272882A (en) | Discrete building detection method and system based on remote sensing image | |
| Manninen et al. | Multi-stage deep learning networks for automated assessment of electricity transmission infrastructure using fly-by images | |
| CN112837281B (en) | Pin defect identification method, device and equipment based on cascade convolution neural network | |
| CN111898427A (en) | Multispectral pedestrian detection method based on feature fusion deep neural network | |
| CN116934696A (en) | Industrial PCB defect detection method and device based on YOLOv7-Tiny model improvement | |
| Wang et al. | High-Voltage Transmission Line Foreign Object and Power Component Defect Detection Based on Improved YOLOv5 | |
| Wu et al. | Ground-based vision cloud image classification based on extreme learning machine | |
| CN116385950A (en) | Electric power line hidden danger target detection method under small sample condition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |