CN115631154A - Power equipment state monitoring and analyzing method and system - Google Patents
Power equipment state monitoring and analyzing method and system Download PDFInfo
- Publication number
- CN115631154A CN115631154A CN202211260089.6A CN202211260089A CN115631154A CN 115631154 A CN115631154 A CN 115631154A CN 202211260089 A CN202211260089 A CN 202211260089A CN 115631154 A CN115631154 A CN 115631154A
- Authority
- CN
- China
- Prior art keywords
- power equipment
- model
- monitoring
- module
- domain
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Abstract
A power equipment state monitoring and analyzing method comprises the following steps: acquiring a defect data set; inputting the defect data set into a transfer learning algorithm, and obtaining a teacher model through learning; inputting the teacher model into a knowledge distillation algorithm to obtain a student model; a student model is deployed on the power equipment to monitor the power equipment. The defect pictures in other scenes are creatively combined with the transfer learning algorithm, the classification of the defects in the power equipment scene is realized, and the problems that the labor intensity is high, the safety is low, the inspection efficiency is low, the defects cannot be found timely and comprehensively and the like in the traditional manual inspection mode are solved.
Description
Technical Field
The invention belongs to the field of deep learning, and particularly relates to a method and a system for monitoring and analyzing the state of power equipment.
Background
Deep learning is widely introduced in the field of monitoring the state of electrical equipment as one of the main branches of machine learning. At present, routing inspection of power equipment of a transformer substation is mainly completed manually, and the traditional manual routing inspection mode has the problems of hard working mode, high labor intensity, low safety, low routing inspection efficiency, untimely defect discovery and the like. Based on power equipment like the semi-manual mode of patrolling and examining of means such as unmanned aerial vehicle, robot, control, most are just stop in the collection function of substation equipment image, and the fault diagnosis of equipment relies on later stage manual identification mostly. Meanwhile, the acquired data needs to be screened, and finally, little and few available data are available. The method not only needs to invest a large amount of manpower and material resources, but also has relatively low diagnosis efficiency. These defects have all greatly restricted the use widely of transformer substation's power equipment inspection platform based on power equipment.
The state of the electric power equipment is monitored and analyzed by using a transfer learning algorithm, and for the problem of insufficient capacity of a sample of the defect of unknown equipment, because some similarities exist between the monitoring data of the defect of the unknown equipment and the monitoring data of the defect of the known equipment, the similarities can be reflected by the features extracted by the correlation algorithm, and therefore the defect features of the unknown equipment can be correctly classified in the algorithm. While a complete detection algorithm based on the transfer learning obtains good performance, the required hardware resources are higher and higher, and the detection algorithm may not be operated on a mobile terminal or an embedded device. A lightweight student network with the same or better performance as a complex transfer learning detection model is trained through knowledge distillation, and can be conveniently carried to the existing power equipment.
Therefore, while the defect detection of the power equipment needs to be performed by using the transfer learning algorithm, the complex neural network model needs to be designed in a light weight manner according to the requirements of limited computing power and real-time performance of the power equipment.
Disclosure of Invention
In order to solve the defects in the prior art, the invention aims to provide a method and a system for monitoring and analyzing the state of electric power equipment, aiming at the problems of the electric power equipment in order to carry out the relevant work of analyzing the defects of the electric power equipment. In the power equipment state monitoring and analyzing scene, the classification of the unknown power equipment defects is realized by adopting a transfer learning technology. And (3) carrying out lightweight design on the obtained complex power equipment defect detection model based on transfer learning by adopting a knowledge distillation technology, and applying the model to various different power equipment.
The invention adopts the following technical scheme.
A power equipment state monitoring and analyzing method comprises the following steps:
s1, acquiring a source domain defect data set with a label and a target domain defect data set without the label.
And S2, constructing a training set and a testing set.
S3, inputting the training set into a transfer learning algorithm, and obtaining a teacher model for monitoring and analyzing the state of the power equipment through iterative training;
s4, inputting the teacher model into a knowledge distillation algorithm to obtain a student model;
and S5, deploying the student model on the electric power equipment to monitor the electric power equipment.
Further, the defect data set is obtained via the internet.
Further, step S1 further includes:
and (3) adopting a translation, rotation, mirror image and Gaussian noise mode for the samples in the defect data set to expand the defect data set.
Further, step S2 specifically includes:
the training set comprises source domain image data with labels and target domain image data without labels, and iterative training is carried out on the teacher model monitored and analyzed by the power equipment by using the training set to obtain the teacher model. The source domain image dataset is an image set of a previous task scene, and the target domain image dataset is an image set of a current task scene (a task scene needing power equipment state monitoring and analysis).
Further, the defect data set is a picture, and step S3 specifically includes:
step S31, acquiring an RGB three-dimensional matrix of a defect data set;
and S32, substituting the RGB three-dimensional matrix into a transfer learning algorithm to obtain a teacher model.
The source domain image and the target domain image are both used as the training set sample, the defect detection model obtained based on the source domain image data set training is directly used for training, the knowledge transfer from the source domain to the target domain is realized, the teacher model obtained after training can adaptively monitor the state of the power equipment on the target domain image, good defect detection precision is obtained in the target domain, the training set does not need to be constructed by labeling the target domain image, the deep learning neural network structure does not need to be reconstructed, and the training time is saved.
Further, step S3 further includes:
the transfer learning algorithm is a sample weight transfer method, a feature transformation transfer method or a model pre-training transfer method. The network structure of the power equipment state monitoring analysis is Faster R-CNN, and the target loss function in the iterative training comprises a confrontation domain adaptive loss and/or a distribution difference loss;
the countermeasure domain adaptive loss is used for representing the domain category explicit of a feature map of a source domain image and a feature map of a target domain image extracted by a feature extractor of a teacher model;
the distribution difference loss is used for representing the distribution difference of candidate region characteristics of the source domain image and candidate region characteristics of the target domain image output by the ROI-pooling network of the teacher model.
Further, the method for acquiring the adaptive loss of the immunity domain comprises the following steps: constructing a domain discriminator network, wherein the domain discriminator network is used for judging whether the feature graph output by the feature extractor is from a source domain or a target domain; the output result according to the domain discriminator network is as followsFormula calculation of adaptive loss L of countermeasure field ad Comprises the following steps:
wherein D is i Representing the output result of the domain discriminator network D corresponding to the ith sample in the training set; domain i A Domain label representing the ith sample in the training set, if Domain i If =0, the ith sample is the source Domain image, if the Domain is not available i =1, the ith sample is the target domain image.
Further, the distribution difference loses L mmd Comprises the following steps:
wherein n is s The number of images representing the source domain image dataset, j and j 'each represent an index of an image in the target domain image dataset, j ≠ j',
representing the jth source domain image,
represents the jth source domain image; n is a radical of an alkyl radical t The number of images representing the target domain image dataset, k and k 'both represent indices of images in the source domain image dataset, k ≠ k',
representing the k-th target domain image,
representing the kth target domain image;
j-th source domain image representing ROI-pooled network output
The candidate region feature of (1);
j' th source domain image representing ROI-pooled network output
The candidate region feature of (1);
k-th target domain image representing ROI-pooled network output
The candidate region feature of (1);
k' th target domain image representing ROI-pooled network output
The candidate region feature of (1); k (a, B) represents a feature kernel for obtaining the candidate region feature a and the candidate region feature B.
Further, in the iterative training, the network parameter of the fast RCNN is continuously adjusted by taking the minimum function value of the target loss function as an optimization target, wherein the target loss function is as follows:
L＝L RPN +L roi +λ(L ad +L mmd )
wherein, represents L RPN Network loss; represents L roi Pooling network loss; λ represents a first weight, λ ∈ [0,3 ]]；L ad Represents the loss of accommodation of the countermeasure domain; l is mmd Indicating a loss of distribution variance.
Further, the defect data set is a picture, and step S4 specifically includes:
the teacher model is the network structure of Resnet50, and the student model is the network structure of Resnet 18; the network structure of Resnet50 or the network structure formula of Resnet50 is:
x l+1 ＝x 1 +F(x l ，w 1 )
where x represents a residual block in the residual network and l represents the number of layers.
Further, the defect data set is a picture, and step S4 specifically includes:
s51, monitoring the power equipment in real time, and extracting data of the power equipment;
and S52, transmitting the data of the power equipment into a student model, carrying out feature extraction and confidence calculation on the data of the power equipment by the student model, and taking the maximum confidence value as the defect category of the current data.
A power device condition monitoring analysis system, comprising: the system comprises a searching module, an algorithm model module and a monitoring module;
the searching module is used for acquiring a defect data set;
the algorithm model module is used for obtaining a teacher model and a student module;
the monitoring module is used for monitoring the power equipment. A
Further, the system further comprises: and the data enhancement module is used for expanding the defect data set.
Further, the algorithm model module comprises: a transfer learning algorithm module and a knowledge distillation algorithm module;
the transfer learning algorithm module is used for obtaining a teacher model;
the knowledge distillation algorithm module is used for obtaining the student module.
Further, the monitoring module comprises: the device comprises a feature extraction module and a calculation module;
the characteristic extraction module is used for extracting the characteristics of the data of the power equipment;
the calculation module is used for confidence calculation.
Compared with the prior art, the invention has the advantages that:
(1) In the power equipment state monitoring and analyzing scene, the number of the scene photos is too small due to the defects of the power equipment, and the principle of confidentiality is adopted. The defect photo and the transfer learning algorithm under other scenes are creatively combined, the defect detection under the power equipment scene is realized, and the problems that the labor intensity is high, the safety is low, the inspection efficiency is low, the defects cannot be found timely and comprehensively in the traditional manual inspection mode are solved.
(2) By adopting a knowledge distillation technology, the obtained complex power equipment defect detection model based on the transfer learning is designed in a light weight manner, a student model with small performance difference with the transfer learning detection model can be optimized, the time consumption is short, the performance is good, and the method is suitable for deployment and prediction.
Drawings
Fig. 1 is a flowchart of a method for monitoring and analyzing a state of an electrical device according to an embodiment of the present invention.
Detailed Description
The present application is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present application is not limited thereby.
Fig. 1 is a flowchart of a method for monitoring and analyzing a state of an electrical device according to an embodiment of the present invention. Aiming at the problems of the power equipment, the embodiment of the invention adopts the technology of combining the transfer learning and knowledge distillation in the state monitoring and analyzing scene of the power equipment to carry out lightweight design on the obtained complex power equipment defect detection model based on the transfer learning so as to realize defect classification.
As shown in fig. 1, the process includes the following steps:
step S1, defect data set of electric power equipment
Further, data enhancement may be performed on the defect data set: and (3) adopting a translation, rotation, mirror image and Gaussian noise mode for the samples in the defect data set to expand the number of the samples in the defect data set.
The defect data set refers to a picture set of a defective power device, and the picture of the defective power device may be a rusted transformer picture, or a picture of some power devices with metal corrosion, oil leakage, or faulty reading. The defect data set may be obtained from an internet search or may be obtained internally from the power station. Secondly, the data enhancement may refer to: and (3) converting a rusted transformer picture by rotation, mirroring, noise addition and the like to obtain more pictures. Therefore, the purpose of the data enhancement is to manually expand the sample, improve the robustness of the model and reduce the risk of overfitting. Specifically, the data set is subjected to data enhancement in the modes of translation, rotation, mirror image and Gaussian noise. In some embodiments, expanding the samples may be rotating a picture by different angles, mirroring, adding noise, changing brightness, etc. to change the picture into multiple pictures. For example, if one picture is rotated by 90 °, 180 ° and mirrored, the picture is expanded by 3 pictures, so that the amount of data is increased.
And S2, constructing a training set. And the training set comprises source domain image data with labels and target domain image data without labels, and the teacher model of the monitoring analysis of the power equipment is subjected to iterative training by using the training set to obtain the teacher model. The source domain image dataset is an image set of a previous task scene, and the target domain image dataset is an image set of a current task scene (a task scene needing power equipment state monitoring and analysis).
And S3, training the enhanced defect data set to obtain a migration learning power equipment defect detection model.
It should be noted that, for example, a defect data set includes several defect pictures, which are unlabeled, i.e., unless we manually click on the confirmation one by one, it is not known what kind of device defect is in the pictures. The migration learning power equipment defect detection model achieves the following purposes: each defect is placed in their respective defect folder. For example: and automatically putting all pictures for indicating the damage into the folders for indicating the damage category, and classifying the pictures for indicating the metal corrosion of the transformer into the folders for indicating the metal corrosion category.
Further, the method for acquiring the adaptive loss of the immunity domain comprises the following steps: constructing a domain discriminator network for discriminating the feature extractor outputWhether the feature map of (a) is derived from a source domain or a target domain; calculating the adaptive loss L of the anti-domain according to the output result of the domain discriminator network according to the following formula ad Comprises the following steps:
wherein D is i Representing the output result of the domain discriminator network D corresponding to the ith sample in the training set; domain i A Domain label representing the ith sample in the training set, if Domain i If =0, the ith sample is the source Domain image, if the Domain is not available i =1, the ith sample is the target domain image.
In the present embodiment, the domain discriminator network D is preferably, but not limited to, a two-class neural network. In the training set training process, a domain discriminator network D is synchronously trained while a front part training sample is used for training a network structure of a crude oil leakage detection model, so that the domain discriminator network D can identify whether a feature map output by a feature extractor is from a source domain image or a target domain image, in the training process of a rear part training sample, the adaptation loss of an anti-domain is calculated, the network parameters of the feature extractor are adjusted by using the adaptation loss of the anti-domain, the domain type of the feature map output by the feature extractor is not obvious, and the domain discriminator network D can not effectively distinguish the source domain image from the target domain image.
In this embodiment, further, the distribution difference loss L mmd Comprises the following steps:
wherein n is s The number of images representing the source domain image dataset, j and j 'each represent an index of an image in the target domain image dataset, j ≠ j',
representing the jth source domain image,
represents the jth source domain image; n is t The number of images representing the target domain image dataset, k and k 'both represent indices of images in the source domain image dataset, k ≠ k',
representing the k-th image of the target field,
representing the kth target domain image;
j-th source domain image representing ROI-pooled network output
The candidate region feature of (1);
j' th source domain image representing ROI-pooled network output
The candidate region feature of (1);
k-th target domain image representing ROI-pooled network output
The candidate region characteristics of (3);
k' th target domain image representing ROI-pooled network output
The candidate region feature of (1); k (a, B) represents a feature kernel for obtaining the candidate region feature a and the candidate region feature B.
Specifically, in the iterative training, the minimum difference between the source domain image features and the target domain image features extracted by the power equipment defect detection model is used as an optimization target to continuously optimize the network parameters of the power equipment defect detection model, so that the knowledge transfer from the source domain to the target domain is realized. The stopping condition of the iterative training is preferably, but not limited to, a preset maximum iteration number, and when the iteration number reaches the maximum iteration number, the iteration is stopped.
Further, the data in the step S2 may be classified by using a transfer learning algorithm, so as to obtain a successfully trained transfer learning power equipment defect detection model. In some embodiments, for example, the defect data set may be a picture, and the nature of the picture is an RGB three-dimensional matrix, and the RGB three-dimensional matrix may be processed by a migration learning algorithm to obtain a one-dimensional matrix, that is, a migration learning power equipment defect detection model, and the value in the matrix is the confidence of the classification result. In the subsequent steps, the highest confidence in the matrix is taken as output, namely the category corresponding to the picture.
It should be noted that the transfer learning algorithm is not the invention of the present application, and all the transfer learning algorithms capable of classifying can be applied to the present solution, and the specific use of which transfer learning algorithm is a routine choice performed by those skilled in the art as needed. The transfer learning algorithm can be any one of a sample weight transfer method, a feature transformation transfer method and a model pre-training transfer method. The transfer learning is to transfer the trained model parameters to a new model to help the training of the new model. Considering that most data are relevant, the learned model parameters (which can also be understood as the knowledge learned by the model) can be shared with the new model through the transfer learning, so that the learning efficiency of the model is accelerated and optimized, and the model does not need to be learned from zero like most networks.
In summary, the migration learning algorithm includes two domains, wherein we call the known domain as the source domain and the domain to be learned as the target domain. In this patent, model parameters obtained by training defect data found on the internet are migrated to defect data obtained inside the power station, so that classification of the defect data inside the power station is realized. That is, the data acquired by the internet is the source domain, and the power equipment defect data is the target domain. The reason for this is that public data sets (which are not all equipment defects but are just finding data similar to defect data as much as possible) are easily obtained from the internet, but real power equipment defect data are difficult to obtain due to the field or the need for confidentiality. Therefore, model parameters trained from easily available data are migrated to data that are difficult to obtain. Certainly, the defect data of the power equipment is difficult to obtain and the quantity is small, so that the method is expanded to obtain more data for model training.
S4, taking the migration learning power equipment defect detection model as a teacher model to train a student model to obtain an optimized student model; description of the invention: the knowledge distillation may select any one of a confrontational distillation algorithm, a multi-teacher knowledge distillation algorithm, a response-based knowledge distillation algorithm, and the like.
Specifically, since the defect detection model is deployed on routing inspection equipment with low computational power, a knowledge distillation technology is used for compressing a generated model (a teacher model) to obtain a small model (a student model) with a relatively simple network structure and low requirement on the computational power of the equipment. The specific operation is as follows: and (3) taking the model after the transfer learning pre-training as a teacher model, and taking the small model without the pre-training (namely the training set in the step S2) as a student model to be put into a knowledge distillation algorithm together for retraining learning. The teacher model adopts a Resnet50 network structure, and the student model adopts a relatively simple Resnet18 network structure, and the difference is that the student model has fewer convolution layers, full connection layers and the like than the teacher model, so the calculation requirement is low, and the obtained student model can be regarded as a newly generated defect detection model.
The formulation of the residual network structure (e.g., the network structure of Resnet50 or the network structure of Resnet 18): x is a radical of a fluorine atom l+1 ＝x l +F(x l ，w l ) Wherein the residual error network structure is composed of a series of residual error blocks, and x represents a residual error block in the residual error networkAnd l represents the number of layers. By recursion, the expression of the characteristics of any deep unit L can be obtained:
feature x for arbitrarily deep cell L L Can be expressed as a feature x of the shallow cell l l Is added with a shape like
Indicating that there is residual behavior between any of the units L and L.
It should be noted that the optimized student model is a model that can be loaded on a device with weak computing power, the number of neural network layers is less than or equal to that of the teacher model, unnecessary parameters in the teacher network model are abandoned, and only necessary parameters are learned from the teacher to identify the detected object. The student model obtains more information by using softmax (soft label), has better generalization capability compared with a teacher model, and can ensure that the identification accuracy of the student model is not inferior to that of the teacher model (original large model) by combining the group route of the original data and the identification of the teacher model although the number of network layers is small. The concrete network structure of the student model is still viewed according to the teacher model.
And S5, deploying the student model on the electric power equipment to monitor the electric power equipment. The method specifically comprises the following steps:
s51, monitoring the power equipment in real time, and extracting data of the power equipment;
step S52, transmitting the data of the power equipment into the defect detection model, carrying out feature extraction and confidence coefficient calculation on the data of the power equipment by the defect detection model, and taking the maximum value of the confidence coefficient as the defect category of the current data; it should be noted that the defect detection model is a student model.
Partial release of the solution to transfer learning is at the very end of the patent.
In summary, it can be understood that, in step S2, a plurality of photographs of a plurality of defect types are input simultaneously, and through the learning in step S3 and step S4, a defect detection model is obtained, and the model can identify all defects. Calculating data of defect detection equipment through a current defect detection model, setting a threshold value of defect confidence coefficient as z0, and if the calculated confidence coefficient of a certain type of defects is greater than z0, considering the classification result of the current data as the type. For example, the confidence coefficients of falling of people, smoking of people and smoking of equipment in the current picture are respectively calculated to be z1, z2 and z3, and according to the definition of the model, the confidence coefficient of z1+ z2+ z3+ defect-free =1. And comparing the sizes of the defects with z0, and assuming that z1 exceeds z0, considering that the current defect is that the person represented by z1 falls. Therefore, it can be understood that the defect category in step S52 includes a non-defective category.
Correspondingly, this application still discloses a power equipment state monitoring analytic system, includes: the system comprises a searching module, an algorithm model module and a monitoring module;
the searching module is used for acquiring a defect data set;
the algorithm model module is used for obtaining a teacher model and a student module;
the monitoring module is used for monitoring the power equipment.
Further, the system further comprises: and the data enhancement module is used for expanding the defect data set.
Further, the algorithm model module comprises: a transfer learning algorithm module and a knowledge distillation algorithm module;
the transfer learning algorithm module is used for obtaining a teacher model;
the knowledge distillation algorithm module is used for obtaining the student module.
Further, the monitoring module comprises: the device comprises a feature extraction module and a calculation module;
the characteristic extraction module is used for extracting the characteristics of the data of the power equipment;
the calculation module is used for confidence calculation.
In summary, the embodiments of the present invention provide a method for monitoring and analyzing a state of an electrical device, which uses a transfer learning technology to classify defects of unknown electrical devices, and solves the problems of high labor intensity, low security, low inspection efficiency, untimely and incomplete defect discovery, and the like in the conventional manual inspection method. By adopting a knowledge distillation technology, the obtained complex power equipment defect detection model based on the transfer learning is designed in a light weight manner, a student model with small performance difference with the transfer learning defect detection model can be optimized, the time consumption is short, the performance is good, and the method is suitable for deployment and prediction.
It should be noted that: the precedence order of the above embodiments of the present invention is only for description, and does not represent the merits of the embodiments. And specific embodiments thereof have been described above. Other embodiments are within the scope of the following claims. In some cases, the steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The present applicant has described and illustrated embodiments of the present invention in detail with reference to the accompanying drawings, but it should be understood by those skilled in the art that the above embodiments are merely preferred embodiments of the present invention, and the detailed description is only for the purpose of helping the reader to better understand the spirit of the present invention, and not for limiting the scope of the present invention, and on the contrary, any improvement or modification made based on the spirit of the present invention should fall within the scope of the present invention.
Claims (14)
1. A power equipment state monitoring and analyzing method is characterized by comprising the following steps:
s1, acquiring a source domain defect data set with a label and a target domain defect data set without the label.
And S2, constructing a training set and a test set.
S3, inputting the training set into a transfer learning algorithm, and obtaining a teacher model for monitoring and analyzing the state of the power equipment through iterative training;
s4, inputting the teacher model into a knowledge distillation algorithm to obtain a student model;
and S5, deploying the student model on the electric power equipment to monitor the electric power equipment.
2. The method for monitoring and analyzing the state of the power equipment according to claim 1, wherein the step S1 further comprises:
and (3) adopting a translation, rotation, mirror image and Gaussian noise mode for the samples in the defect data set to expand the defect data set.
3. The method for monitoring and analyzing the state of the power equipment according to claim 1, wherein the step S2 further comprises:
the defect data set is divided into a training set and a test set. The training set includes source domain image data with labels and target domain image data without labels. The test set includes target domain image data without a label.
4. The method for monitoring and analyzing the state of the power equipment according to claim 1, wherein the defect data set is a picture, and the step S3 specifically includes:
step S31, acquiring an RGB three-dimensional matrix of a defect data set;
and S32, substituting the RGB three-dimensional matrix into a transfer learning algorithm to obtain a teacher model.
5. The power equipment state monitoring and analyzing method according to claim 1, wherein the transfer learning algorithm is a sample weight transfer method, a feature transformation transfer method or a model pre-training transfer method. The network structure of the power equipment state monitoring analysis is Faster R-CNN, and the target loss function in the iterative training comprises a confrontation domain adaptive loss and/or a distribution difference loss;
the countermeasure domain adaptive loss is used for representing the domain category explicit of a feature map of a source domain image and a feature map of a target domain image extracted by a feature extractor of a teacher model;
the distribution difference loss is used for representing the distribution difference of candidate region characteristics of the source domain image and the candidate region characteristics of the target domain image output by the ROI-pooling network of the teacher model.
6. The power equipment state monitoring analysis method according to claim 5, wherein the method for acquiring the domain adaptation loss comprises the following steps:
constructing a domain discriminator network, wherein the domain discriminator network is used for judging whether the feature graph output by the feature extractor is from a source domain or a target domain; calculating the adaptive loss L of the anti-domain according to the output result of the domain discriminator network according to the following formula ad Comprises the following steps:
wherein D is i Representing the output result of the domain discriminator network D corresponding to the ith sample in the training set; domain i A Domain label representing the ith sample in the training set, if Domain i If =0, the ith sample is the source Domain image, if the Domain is not available i =1, the ith sample is the target domain image.
7. A power equipment condition monitoring and analyzing method according to claim 2 or 3, wherein said distribution difference loss L mmd Comprises the following steps:
wherein n is s The number of images representing the source domain image dataset, j and j 'each represent an index of an image in the target domain image dataset, j ≠ j',
representing the jth source domain image,
represents the jth source domain image; n is t The number of images representing the target domain image dataset, k and k 'both represent indices of images in the source domain image dataset, k ≠ k',
representing the k-th target domain image,
representing the kth target domain image;
j-th source domain image representing ROI-pooled network output
The candidate region feature of (1);
j' th source domain image representing ROI-pooled network output
The candidate region feature of (1);
k-th target domain image representing ROI-pooled network output
The candidate region feature of (1);
k' th target domain image representing ROI-pooled network output
The candidate region feature of (1); k (A, B) represents the feature of obtaining the candidate region feature A and the candidate region feature BAnd (5) checking the kernel.
8. The method according to claim 2 or 3, wherein the iterative training continuously adjusts the network parameters of the fast RCNN with the minimum function value of the objective loss function as the optimization target, and the objective loss function is:
L＝L RPN +L roi +λ(L ad +L mmd )
wherein, represents L RPN Network loss; represents L roi Pooling network loss; λ represents a first weight, λ ∈ [0,3 ∈ [ ]]；L ad Represents the loss of accommodation of the countermeasure domain; l is mmd Indicating a loss of distribution variance.
9. The power equipment state monitoring and analyzing method of claim 1, wherein the knowledge distillation algorithm is a countermeasure distillation algorithm, a multi-teacher knowledge distillation algorithm, or a response-based knowledge distillation algorithm.
10. The method for monitoring and analyzing the state of the power equipment according to claim 1, wherein the defect data set is a picture, and the step S4 specifically comprises:
the teacher model is the network structure of Resnet50, and the student model is the network structure of Resnet 18;
the network structure of Resnet50 or the network structure formula of Resnet50 is:
x l+1 ＝x l +F(x l ,w l )
where x represents a residual block in the residual network and l represents the number of layers.
11. The method for monitoring and analyzing the state of the power equipment according to claim 1, wherein the defect data set is a picture, and the step S5 specifically comprises:
s51, monitoring the power equipment in real time, and extracting data of the power equipment;
and S52, transmitting the data of the power equipment into a student model, carrying out feature extraction and confidence calculation on the data of the power equipment by the student model, and taking the maximum confidence value as the defect category of the current data.
12. A power equipment condition monitoring and analysis system for performing the method of claims 1-11, the system comprising: the system comprises a searching module, an algorithm model module and a monitoring module;
the searching module is used for acquiring a defect data set;
the algorithm model module is used for obtaining a teacher model and a student model;
the monitoring module is used for monitoring the power equipment.
13. The system for monitoring and analyzing the condition of the power equipment according to claim 11, wherein the system further comprises: and the data enhancement module is used for expanding the defect data set.
14. The system of claim 11, wherein the algorithmic model module comprises: a transfer learning algorithm module and a knowledge distillation algorithm module;
the transfer learning algorithm module is used for obtaining a teacher model;
the knowledge distillation algorithm module is used for obtaining the student module.
The monitoring module includes: the characteristic extraction module and the calculation module;
the characteristic extraction module is used for extracting the characteristics of the data of the power equipment;
the calculation module is used for confidence calculation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211260089.6A CN115631154A (en) | 2022-10-14 | 2022-10-14 | Power equipment state monitoring and analyzing method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211260089.6A CN115631154A (en) | 2022-10-14 | 2022-10-14 | Power equipment state monitoring and analyzing method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115631154A true CN115631154A (en) | 2023-01-20 |
Family
ID=84904755
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211260089.6A Pending CN115631154A (en) | 2022-10-14 | 2022-10-14 | Power equipment state monitoring and analyzing method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115631154A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117274723A (en) * | 2023-11-22 | 2023-12-22 | 国网智能科技股份有限公司 | Target identification method, system, medium and equipment for power transmission inspection |
-
2022
- 2022-10-14 CN CN202211260089.6A patent/CN115631154A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117274723A (en) * | 2023-11-22 | 2023-12-22 | 国网智能科技股份有限公司 | Target identification method, system, medium and equipment for power transmission inspection |
| CN117274723B (en) * | 2023-11-22 | 2024-03-26 | 国网智能科技股份有限公司 | Target identification method, system, medium and equipment for power transmission inspection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107944396B (en) | Knife switch state identification method based on improved deep learning | |
| WO2021046951A1 (en) | Image identification method, system, and storage medium | |
| CN111563557B (en) | Method for detecting target in power cable tunnel | |
| CN111179249A (en) | Power equipment detection method and device based on deep convolutional neural network | |
| CN110717526A (en) | Unsupervised transfer learning method based on graph convolution network | |
| CN110874590B (en) | Training and visible light infrared visual tracking method based on adapter mutual learning model | |
| CN111524132B (en) | Method, device and storage medium for identifying abnormal cells in sample to be detected | |
| CN111860106B (en) | Unsupervised bridge crack identification method | |
| CN115294117B (en) | Defect detection method and related device for LED lamp beads | |
| CN112633382A (en) | Mutual-neighbor-based few-sample image classification method and system | |
| CN111223087B (en) | Automatic bridge crack detection method based on generation countermeasure network | |
| WO2024021461A1 (en) | Defect detection method and apparatus, device, and storage medium | |
| CN113436184A (en) | Power equipment image defect judging method and system based on improved twin network | |
| CN116385958A (en) | Edge intelligent detection method for power grid inspection and monitoring | |
| CN115631154A (en) | Power equipment state monitoring and analyzing method and system | |
| CN113837994B (en) | Photovoltaic panel defect diagnosis method based on edge detection convolutional neural network | |
| Saleh et al. | Tire defect detection model using machine learning | |
| CN111210412A (en) | Package detection method and device, electronic equipment and storage medium | |
| Prihatno et al. | Metal defect classification using deep learning | |
| CN112419243A (en) | Power distribution room equipment fault identification method based on infrared image analysis | |
| CN117036904A (en) | Attention-guided semi-supervised corn hyperspectral image data expansion method | |
| CN116681962A (en) | Power equipment thermal image detection method and system based on improved YOLOv5 | |
| CN115393802A (en) | Railway scene unusual invasion target identification method based on small sample learning | |
| CN111274986B (en) | Dish identification and classification method based on image analysis | |
| Shishkin et al. | Implementation of yolov5 for detection and classification of microplastics and microorganisms in marine environment |
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 |