CN117274748B - Lifelong learning power model training and detecting method based on outlier rejection - Google Patents
Lifelong learning power model training and detecting method based on outlier rejection Download PDFInfo
- Publication number
- CN117274748B CN117274748B CN202311526946.7A CN202311526946A CN117274748B CN 117274748 B CN117274748 B CN 117274748B CN 202311526946 A CN202311526946 A CN 202311526946A CN 117274748 B CN117274748 B CN 117274748B
- Authority
- CN
- China
- Prior art keywords
- power
- training
- defect detection
- representing
- detection model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000012549 training Methods 0.000 title claims abstract description 105
- 238000000034 method Methods 0.000 title claims abstract description 66
- 230000007547 defect Effects 0.000 claims abstract description 163
- 238000001514 detection method Methods 0.000 claims abstract description 144
- 230000005540 biological transmission Effects 0.000 claims abstract description 38
- 230000008569 process Effects 0.000 claims abstract description 31
- 230000004927 fusion Effects 0.000 claims abstract description 20
- 230000001629 suppression Effects 0.000 claims abstract description 10
- 230000006870 function Effects 0.000 claims description 33
- 230000014509 gene expression Effects 0.000 claims description 26
- 238000011176 pooling Methods 0.000 claims description 19
- 238000010586 diagram Methods 0.000 claims description 14
- 238000013507 mapping Methods 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 7
- 230000003213 activating effect Effects 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 210000000988 bone and bone Anatomy 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims 1
- 230000002159 abnormal effect Effects 0.000 abstract description 8
- 230000008014 freezing Effects 0.000 abstract description 5
- 238000007710 freezing Methods 0.000 abstract description 5
- 230000036961 partial effect Effects 0.000 abstract description 4
- 230000009286 beneficial effect Effects 0.000 description 13
- 238000004364 calculation method Methods 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000005299 abrasion Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
Classifications
-
- 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
- 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
- G06N3/048—Activation functions
-
- 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
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Medical Informatics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Biodiversity & Conservation Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a lifetime learning power model training and detecting method based on outlier rejection, which comprises the following steps: acquiring power defect image data of a power transmission and distribution line; training to obtain an electric power defect detection model; performing self-adaptive suppression on outlier data of a characteristic pyramid module in the electric power defect detection model through a dense region fusion algorithm, and optimizing to obtain a final trained electric power defect detection model; and performing defect detection on the power transmission and distribution line power defect image data by using the trained power defect detection model. According to the invention, through continuous incremental training of the electric power defect image data in the lifetime learning process and self-adaptive suppression of outlier data, abnormal parameters in the model are effectively suppressed, the problem of abnormal parameters caused by full-scale updating in the existing training method and the problem of incomplete sample feature learning caused by partial parameter freezing updating are solved, and the detection precision is further improved.
Description
Technical Field
The invention belongs to the field of power defect detection model training, and particularly relates to a lifetime learning power model training and detection method based on outlier rejection.
Background
In recent years, with the increasing demand of electric power in China, an electric power system has become one of the most indispensable infrastructures in people's production and life. As an important component of the power system, the power transmission and distribution line plays an important role in power transmission, and maintenance and management of the power transmission and distribution line have important significance for stable operation of the power system. Because power transmission and distribution lines are often subjected to abrasion and influence of various environments and external factors, various potential safety hazards and defects exist. Therefore, finding and repairing these defects in time is particularly important for power systems.
Compared with manual detection, the machine learning method improves defect detection efficiency. However, the traditional offline machine learning method has the problems of outdated training data, poor model generalization capability and the like, and particularly under the scene of continuous attention and maintenance of a power system, the intelligent degree and instantaneity of the model cannot meet the actual needs. The lifelong learning training method which is rising in recent years can well solve the problems of the traditional machine learning method, and can realize gradually intelligent tasks and functions in the process of continuously updating and continuously expanding the model.
The existing life-long learning electric power defect detection model training method is mainly divided into two types, wherein the model is finely adjusted by adopting full-quantity parameters. However, the method has the defects that the incremental training process generates abnormal parameters and damages the performance of the model because more parameters need to be updated and differences between the previous parameters and the updated parameters are not considered when the model is updated. Another type is to use freezing of part of the parameters to fine tune the model. However, since all parameters of the model are not updated, the model cannot completely learn the latest updated sample characteristics.
Fine tuning of the model during life-time learning, using full-scale parameter updates can enable the model to learn more features of new data samples. Meanwhile, model parameters are analyzed, outlier data with negative influence on the model is adaptively restrained, and model parameter differences caused by new and old data iteration can be removed. Therefore, in order to solve the above-mentioned problems in the prior art, a power model training method for lifetime learning of adaptive outlier data suppression is needed to train a power defect detection model.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides the life-long learning power model training and detecting method based on outlier suppression. The method solves the problem of abnormal parameters caused by full-scale updating in the existing training method and the problem of incomplete sample feature learning caused by partial parameter freezing updating, thereby improving the detection precision.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the scheme provides a lifetime learning power model training and detecting method based on outlier rejection, which comprises the following steps:
s1, acquiring power defect image data of a power transmission and distribution line;
s2, in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid by using a lifetime learning model of model fine adjustment to obtain an electric power defect detection model;
s3, performing self-adaptive suppression on outlier data of a feature pyramid module in the electric power defect detection model through a dense region fusion algorithm, and optimizing to obtain a final trained electric power defect detection model;
and S4, performing defect detection on the power transmission and distribution line power defect image data by using the trained power defect detection model.
The beneficial effects of the invention are as follows: the method comprises the steps of establishing a lifetime parameter set for parameters of each position of a feature pyramid module in the electric defect detection model after full parameter fine adjustment, and continuously adding and adjusting the parameter set in the whole lifetime learning process; then, carrying out regional division on the set to obtain a plurality of power parameter regions, carrying out calculation and analysis on the plurality of parameter regions and inhibiting outlier data in the plurality of parameter regions; and finally, acquiring a dense region, averaging the parameters in the region to obtain a fusion parameter, acquiring a final electric power defect detection model, and detecting by using the electric power defect detection model. According to the invention, the abnormal data brought by new and old data in the life learning process is adaptively restrained, so that the electric power defect detection model can be dynamically updated and expanded continuously, and the electric power defect detection model can effectively detect the defects of the power transmission and distribution line for the life. The invention solves the problem of abnormal parameters caused by full-scale updating and incomplete sample feature learning caused by partial parameter freezing updating in the existing model training method, thereby improving the detection precision.
Further, the step S2 includes the steps of:
s201, dividing power transmission and distribution line power defect image data into training data sets;
s202, using a training data set, and in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid network by using a lifetime learning model of model fine adjustment to obtain an electric power defect detection model.
The beneficial effects of the above-mentioned further scheme are: the invention adopts the model fine-tuning lifelong learning model to carry out incremental training in the whole lifelong learning process, thereby retaining the former data characteristics and enabling the fusion model to effectively extract the data information of different characteristics.
Still further, the step S202 includes the steps of:
s2021, if the current training state is the initial training in the model fine-tuning lifetime learning paradigm process, entering S2022, otherwise, entering S2023;
s2022, extracting a back bone network by taking a ResNet50 network as a characteristic, extracting a Neck network by taking a characteristic pyramid network as a characteristic, and constructing a complete electric power defect detection model by combining a classification network and a bounding box regression network;
s2023, calling a power defect detection model used in the previous training in the lifetime learning process;
s2024, training the electric power defect detection model according to a lifetime learning paradigm of model fine adjustment to obtain the electric power defect detection model, wherein the training process comprises the following steps: activating network neurons in the power defect detection model by using a Relu activation function, performing maximum pooling operation on the power characteristic diagram extracted from the power defect detection model, and calculating loss on a classification network and a bounding box regression network in the power defect detection model.
The beneficial effects of the above-mentioned further scheme are: according to the invention, the ResNet50 and the feature pyramid module are selected as the feature extraction network, so that feature extraction can be carried out on sample data from two aspects of depth and breadth, and the electric power defect detection model can classify the types of the data samples and can position the defect positions of the samples by combining the classification network and the bounding box regression network.
Still further, the expression of the lifetime learning paradigm of the model fine tuning is as follows:
wherein,representing life-long learningnA power defect detection model for secondary incremental training; />Representing a power transmission and distribution line defect trained using non-acquired and +.>The power defect detection models with the same structure,representing life-long learningn-1 incrementally trained power defect detection model, < >>Representing life-long learningnTraining in secondary incrementsData set used by the power defect detection model, < >>Represents the number of training increments>Expressed as +.>For training data set, and for the firstn-1 incrementally trained power defect detection model->Continuing training, ->For the indication function, an expression corresponding to the indication function is adopted when the indication function meets the condition.
The beneficial effects of the above-mentioned further scheme are: according to the invention, the electric power defect detection model is trained in a model fine-tuning lifelong learning paradigm increment mode, and high-quality parameter information of each increment model is provided for fusing the electric power defect detection model.
Still further, the expression for activating the network neurons in the power defect detection model is as follows:
wherein,network neurons representing inputs, +.>Network neurons representing outputs, when +.>When the ratio is not more than 0,when->Greater than 0->，/>Representing a maximization operation.
The beneficial effects of the above-mentioned further scheme are: the Relu activation function is adopted, the input signal of the neural network can be subjected to nonlinear mapping, so that the nonlinear modeling capability of the neural network is realized, and in addition, the Relu activation function has the advantages of simplicity in calculation, stability in gradient and the like, and the problem of gradient disappearance can be effectively avoided.
Still further, the expression of the max-pooling operation is as follows:
wherein,representing the first of the power defect detection modelsjThe first power characteristic diagramkThe output values of the individual pooling windows are,representing the first of the power defect detection modelsjThe first power characteristic diagramkThe first area covered by the pooling windowiThe value of the individual pixels>Representing the first of the power defect detection modelsjThe first power characteristic diagramkAreas covered by individual pooling windows, +.>Representing a maximization operation.
The beneficial effects of the above-mentioned further scheme are: the image features are downsampled by maximum pooling, thereby reducing the size of the image and extracting the main features of the image.
Still further, the expressions for calculating the losses of the classification network and the bounding box regression network in the power defect detection model are as follows:
wherein,representing the loss function of the classification network,Na total number of samples representing the power defect image data,indicate->The true value of the individual power defect image data samples, is->Indicate->Predictive value of individual power defect image data samples, < >>Representing a logarithmic function>Respectively representing a set of predicted values and real values;
wherein,loss function representing a regression network of bounding boxes, +.>Representing the actual value +.>Representing the predicted value.
The beneficial effects of the above-mentioned further scheme are: by adopting the cross entropy loss function, the electric power defect detection model can finally reduce the value of the cross entropy loss function between the predicted value and the true value as much as possible through learning and continuously optimizing the training set data, thereby improving the performance of the electric power defect detection model. The invention adopts the Smooth L1 loss function, can reduce the problem of excessive fitting caused by local abnormal values, and improves the accuracy and stability of the regression model. In addition, the Smooth L1 loss function can identify and restrain noise in input data, so that reliability and robustness of the power defect detection model are improved.
Still further, the step S3 includes the steps of:
s301, acquiring a feature pyramid module in the electric power defect detection model, and performing dimension mapping on the feature pyramid module to obtain original dimension information of the feature pyramid module;
s302, judging whether the current electric power defect detection model training is the primary training or the incremental training in the life learning process, if the current electric power defect detection model training is the primary training, entering S303, otherwise, entering S304;
s303, in the feature pyramid moduleNo. 5 of the individual positions>Individual parameters constitute the lifetime parameter set +.>And proceeds to S305;
s304, in the feature pyramid moduleNo. 5 of the individual positions>Personal parameter addition to the final parameter set->And proceeds to S305;
s305, according to the lifetime parameter setDividing the power transmission line into a plurality of power transmission line parameter areas>；
S306, according to the multiple power transmission and distribution line parameter areasBy calculating the quantity and distribution discrete degree of parameters in the region, the outlier data with negative influence on the electric power defect detection model is adaptively restrained, and the electric power dense region is obtained；
S307, according to the power dense regionCalculating to obtain the characteristic pyramid module +.>Electric power fusion parameter in individual locations->；
S308, judging whether all parameters in the feature pyramid module are fused, if so, restoring the feature pyramid module to the original dimension by using the original dimension information of the feature pyramid module, and entering S309, otherwise, returning to the step S302;
s309, combining the feature pyramid module obtained after the processing of S308 with a network module except the feature pyramid module in the electric power defect detection model, and optimizing to obtain the finally trained electric power defect detection model.
The beneficial effects of the above-mentioned further scheme are: according to the training state of the training in the lifetime learning process, the invention utilizes the dense region fusion algorithm to carry out self-adaptive suppression on outlier data in the basic electric power defect detection model feature pyramid module, fuses the defect detection model with stable performance for the training in the lifetime learning process, and simultaneously provides a model foundation for the next incremental training in the lifetime learning process.
Still further, the step S301 includes the steps of:
s3011, acquiring parameter dimensionality of a feature pyramid module in the electric power defect detection modelWherein, the method comprises the steps of, wherein,respectively representing the size of the parameter matrix in three dimensions of length, width and height;
s3012, traversing all parameters according to the parameter dimension by the feature pyramid module, and mapping all traversed parameters to a one-dimensional space through dimension mapping;
s3013, obtaining the original dimension information of the feature pyramid module according to the processing result of S3012.
The beneficial effects of the above-mentioned further scheme are: according to the invention, the feature pyramid network parameters of the electric power defect detection model are mapped to one dimension, so that the feature layers with different sizes in the electric power defect detection model are better processed uniformly; and meanwhile, the original dimension information of the parameters of the feature pyramid module is recorded, so that dimension recovery is facilitated after the parameters are fused.
Still further, the dimension map is calculated as follows:
wherein,representing dimension map->Respectively representing the coordinate values of the parameters in the dimension space.
The beneficial effects of the above-mentioned further scheme are: according to the invention, the dimension mapping is carried out on the characteristic pyramid network parameters of the electric power defect detection model according to the length, width and height dimensions and the coordinate values of the parameters, the formula calculation complexity is low, and the calculation time of the dimension mapping step is reduced.
Still further, the step S305 includes the steps of:
s3051, set of lifetime parametersThe average division of the value range between the maximum and minimum values is +.>A plurality of regions;
s3052, collecting the lifetime parametersAll parameters in the power transmission and distribution line are distributed to corresponding parameter areas according to the values to obtain a plurality of power transmission and distribution line parameter areas +.>。
Still further, the division intoThe expression of the individual regions is as follows:
wherein,and->Respectively representing lifetime parameter sets->Maximum and minimum of (a), j->The number of divided regions is indicated,drepresenting the size of each region after division, +.>Indicate->Left end point of individual region,/>Indicate->The right end point of the region.
The beneficial effects of the above-mentioned further scheme are: for lifetime parameter setsThe regional division is performed, so that the analysis of parameter distribution and parameter change in the whole life learning process is facilitated, and the outlier data generated by new and old data iteration is found.
Still further, the power dense regionThe expression of (2) is as follows:
wherein,representing the +.>Parameter set on individual positions, +.>Representing multiple power transmission and distribution line parameter areas>The number of elements in the collection, +.>Representing variance->Representing a preset threshold value,/->Representing a maximization operation, +_>For indicating function, if->Empty, then->And the value of (2) is 1, otherwise 0.
The beneficial effects of the above-mentioned further scheme are: the variance is calculated through the number of parameters in the parameter area, the dense area is dynamically selected according to the variance, outlier data in other parameter areas are restrained in a self-adaptive mode, and the parameter area with good expression capability on all the power defect image data is obtained.
Still further, the electricityForce fusion parametersThe expression of (2) is as follows:
wherein,nrepresenting power dense areasThe number of elements in the collection, +.>Representing a power dense region->Is a component of the group.
The beneficial effects of the above-mentioned further scheme are: and obtaining fusion parameters with good expression capacity for all the electric power defect image data by carrying out average fusion on the parameters of the electric power dense region.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a diagram of a structure of a dense region fusion algorithm in this embodiment.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and all the inventions which make use of the inventive concept are protected by the spirit and scope of the present invention as defined and defined in the appended claims to those skilled in the art.
Examples
As shown in fig. 1, the invention provides a lifetime learning power model training and detecting method based on outlier rejection, which comprises the following implementation steps:
s1, acquiring power defect image data of a power transmission and distribution line;
s2, in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid by using a lifetime learning model of model fine adjustment to obtain an electric power defect detection model, wherein the implementation method comprises the following steps:
s201, dividing power transmission and distribution line power defect image data into training data sets;
s202, using a training data set, in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid network by using a lifetime learning paradigm of model fine tuning to obtain an electric power defect detection model, wherein the implementation method comprises the following steps:
s2021, if the current training state is the initial training in the model fine-tuning lifetime learning paradigm process, entering S2022, otherwise, entering S2023;
s2022, extracting a back bone network by taking a ResNet50 network as a characteristic, extracting a Neck network by taking a characteristic pyramid network as a characteristic, and constructing a complete electric power defect detection model by combining a classification network and a bounding box regression network;
s2023, calling a power defect detection model used in the previous training in the lifetime learning process;
s2024, training the electric power defect detection model according to a lifetime learning paradigm of model fine adjustment to obtain the electric power defect detection model, wherein the training process comprises the following steps: activating network neurons in the power defect detection model by using a Relu activation function, carrying out maximum pooling operation on the power characteristic diagram extracted from the power defect detection model, and calculating loss on a classification network and a bounding box regression network in the power defect detection model;
the expression of the lifetime learning paradigm of model fine tuning is as follows:
wherein,representing life-long learningnA power defect detection model for secondary incremental training; />Representing a power transmission and distribution line defect trained using non-acquired and +.>The power defect detection models with the same structure,representing life-long learningn-1 incrementally trained power defect detection model, < >>Representing life-long learningnData set used by the power defect detection model for secondary incremental training, < >>Represents the number of training increments>Expressed as +.>For training data set, and for the firstn-1 incrementally trained power defect detection model->Continuing training, ->The method comprises the steps of representing an expression corresponding to an indication function when the indication function meets the condition;
the expression for activating the network neurons in the power defect detection model is as follows:
wherein,network neurons representing inputs, +.>Network neurons representing outputs, when +.>When the ratio is not more than 0,when->Greater than 0->，/>Representing a maximization operation.
The expression of the max pooling operation is as follows:
wherein,representing the first of the power defect detection modelsjThe first power characteristic diagramkThe output values of the individual pooling windows are,representing the first of the power defect detection modelsjThe first power characteristic diagramkThe first area covered by the pooling windowiThe value of the individual pixels>Representing the first of the power defect detection modelsjThe first power characteristic diagramkAreas covered by individual pooling windows, +.>Representing a maximization operation, typically a rectangular region. In maximum pooling, the maximum value is taken as the value of the corresponding position of the feature map output by the pooling layer for all elements in the pooling window.
The expressions for calculating the losses of the classification network and the bounding box regression network in the electric power defect detection model are respectively as follows:
wherein,representing the loss function of the classification network,Na total number of samples representing the power defect image data,indicate->The true value of the individual power defect image data samples, is->Indicate->Predictive value of individual power defect image data samples, < >>Representing a logarithmic function>Respectively representing a set of predicted values and real values;
wherein,representing bounding boxesRegression network loss function +.>Representing the actual value +.>Representing the predicted value. The boundary box regression network loss function adopts a Smooth L1 loss function, when the difference between the actual value and the prediction is smaller than 1, a square term of the L2 loss function is used, and when the difference is larger than or equal to 1, the difference term is the L1 loss function, and 0.5 is subtracted.
S3, performing self-adaptive suppression on outlier data of a feature pyramid module in the electric power defect detection model through a dense region fusion algorithm, optimizing and obtaining a final trained electric power defect detection model, wherein the implementation method comprises the following steps:
s301, acquiring a feature pyramid module in the electric power defect detection model, and performing dimension mapping on the feature pyramid module to obtain original dimension information of the feature pyramid module, wherein the implementation method is as follows:
s3011, acquiring parameter dimensionality of a feature pyramid module in the electric power defect detection modelWherein, the method comprises the steps of, wherein,respectively representing the size of the parameter matrix in three dimensions of length, width and height;
s3012, traversing all parameters according to the parameter dimension by the feature pyramid module, and mapping all traversed parameters to a one-dimensional space through dimension mapping;
the calculation formula of the dimension map is as follows:
wherein,representing dimension map->Respectively representing the coordinate values of the parameters in the dimension space;
s3013, obtaining original dimension information of the feature pyramid module according to the processing result of S3012;
s302, judging whether the current electric power defect detection model training is the primary training or the incremental training in the life learning process, if the current electric power defect detection model training is the primary training, entering S303, otherwise, entering S304;
s303, in the feature pyramid moduleNo. 5 of the individual positions>Individual parameters constitute the lifetime parameter set +.>And proceeds to S305;
s304, in the feature pyramid moduleNo. 5 of the individual positions>Personal parameter addition to the final parameter set->And proceeds to S305;
s305, according to the lifetime parameter setDividing the power transmission line into a plurality of power transmission line parameter areas>The implementation method is as follows:
s3051, set of lifetime parametersThe average division of the value range between the maximum and minimum values is +.>A plurality of regions;
the division intoThe expression of the individual regions is as follows:
wherein,and->Respectively representing lifetime parameter sets->Maximum and minimum of (a), j->The number of divided regions is indicated,drepresenting the size of each region after division, +.>Indicate->Of individual areasLeft end point,/->Indicate->Right end point of each region;
s3052, collecting the lifetime parametersAll parameters in the power transmission and distribution line are distributed to corresponding parameter areas according to the values to obtain a plurality of power transmission and distribution line parameter areas +.>；
S306, according to the multiple power transmission and distribution line parameter areasBy calculating the quantity and distribution discrete degree of parameters in the region, the outlier data with negative influence on the electric power defect detection model is adaptively restrained, and the electric power dense region is obtained；
The power dense regionThe expression of (2) is as follows:
wherein,representing the +.>Parameter set on individual positions, +.>Representing a plurality ofPower transmission and distribution line parameter area->The number of elements in the collection, +.>Representing variance->Representing a preset threshold value,/->Representing a maximization operation, +_>For indicating function, if->Empty, then->The value of (2) is 1, otherwise 0;
s307, according to the power dense regionCalculating to obtain the characteristic pyramid module +.>Electric power fusion parameter in individual locations->；
The power fusion parametersThe expression of (2) is as follows:
wherein,nrepresenting power dense areasThe number of elements in the collection, +.>Representing a power dense region->Elements of (a) and (b);
s308, judging whether all parameters in the feature pyramid module are fused, if so, restoring the feature pyramid module to the original dimension by using the original dimension information of the feature pyramid module, and entering S309, otherwise, returning to the step S302;
s309, combining the feature pyramid module obtained after the processing of S308 with a network module except the feature pyramid module in the electric power defect detection model, and optimizing to obtain the finally trained electric power defect detection model.
And S4, performing defect detection on the power transmission and distribution line power defect image data by using the trained power defect detection model.
According to the power defect model training method for lifetime learning of self-adaptive outlier data suppression, which is related to the invention, the parameters of the power defect detection model are updated by adopting full parameter fine adjustment in the whole lifetime learning process, so that all the characteristics of new data are learned to the greatest extent. Meanwhile, outlier data generated by new and old data iteration in a feature pyramid module of the electric power defect detection model is also analyzed, and the final electric power defect detection model is obtained by adaptively restraining the outlier data and carrying out average fusion on residual parameters. The structure diagram of the dense region fusion algorithm is shown in fig. 2, and in fig. 2, three steps are included: a data processing stage, a pre-training stage of the electric power defect detection model and an acquisition stage of the electric power defect detection model. D represents an undivided data set,data set representing use of power defect detection model for lifelong learning 1 st increment training,/>Data set representing use of power defect detection model for lifetime learning 2 nd increment training,/, for example>Representing life-long learningnData set used by the power defect detection model for secondary incremental training, < >>Power defect detection model representing life-long learning 1 st incremental training, ++>Indicates the study of the lifetime->Electric defect detection model trained in secondary increment, DRF (dynamic random field) represents specific process of algorithm based on dense region fusion, ++>And representing the electric power defect detection model finally subjected to DRF fusion. Through the scheme, the power defect detection model can continuously carry out self-adaptive suppression on outlier data and continuously adjust and optimize the outlier data in the whole life learning process, so that good defect detection capability is maintained on power transmission and distribution line defect power image data. The method solves the problem of abnormal parameters caused by full-scale updating in the existing power model training method and the problem of incomplete characteristic learning of the power defect data sample caused by partial parameter freezing updating, thereby improving the detection precision. />
Claims (11)
1. The life-long learning power model training and detecting method based on outlier rejection is characterized by comprising the following steps of:
s1, acquiring power defect image data of a power transmission and distribution line;
s2, in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid by using a lifetime learning model of model fine adjustment to obtain an electric power defect detection model;
the expression of the lifetime learning paradigm of model fine tuning is as follows:
wherein,representing life-long learningnA power defect detection model for secondary incremental training; />Representing a power transmission and distribution line defect trained using non-acquired and +.>Electric power defect detection model with same structure +.>Representing life-long learningn-1 incrementally trained power defect detection model, < >>Representing life-long learningnData set used by the power defect detection model for secondary incremental training, < >>Represents the number of training increments>Expressed as +.>For training data set, and for the firstn-1 incrementally trained power defect detection model->Continuing training, ->The method comprises the steps of representing an expression corresponding to an indication function when the indication function meets the condition;
s3, performing self-adaptive suppression on outlier data of a feature pyramid module in the electric power defect detection model through a dense region fusion algorithm, and optimizing to obtain a final trained electric power defect detection model;
the step S3 comprises the following steps:
s301, acquiring a feature pyramid module in the electric power defect detection model, and performing dimension mapping on the feature pyramid module to obtain original dimension information of the feature pyramid module;
s302, judging whether the current electric power defect detection model training is the primary training or the incremental training in the life learning process, if the current electric power defect detection model training is the primary training, entering S303, otherwise, entering S304;
s303, in the feature pyramid moduleNo. 5 of the individual positions>Individual parameters constitute the lifetime parameter set +.>And proceeds to S305;
s304, in the feature pyramid moduleNo. 5 of the individual positions>Personal parameter addition to the final parameter set->And proceeds to S305;
s305, according to the lifetime parameter setDividing the power transmission line into a plurality of power transmission line parameter areas>；
S306, according to the multiple power transmission and distribution line parameter areasBy calculating the quantity and distribution discrete degree of parameters in the region, the outlier data with negative influence on the electric power defect detection model is adaptively restrained, and the electric power dense region +.>；
S307, according to the power dense regionCalculating to obtain the characteristic pyramid module +.>Electric power fusion parameter in individual locations->；
S308, judging whether all parameters in the feature pyramid module are fused, if so, restoring the feature pyramid module to the original dimension by using the original dimension information of the feature pyramid module, and entering S309, otherwise, returning to the step S302;
s309, combining the feature pyramid module obtained after the processing of S308 with a network module except the feature pyramid module in the electric power defect detection model, and optimizing to obtain a final trained electric power defect detection model;
the power dense regionThe expression of (2) is as follows:
wherein,representing the +.>Parameter set on individual positions, +.>Representing multiple power transmission and distribution line parameter areas>The number of elements in the collection, +.>Representing variance->Representing a preset threshold value,/->Representing a maximization operation, +_>For indicating function, if->Empty, then->The value of (2) is 1, otherwise 0;
and S4, performing defect detection on the power transmission and distribution line power defect image data by using the trained power defect detection model.
2. The method for training and detecting a lifetime learning power model based on outlier rejection according to claim 1, wherein S2 comprises the steps of:
s201, dividing power transmission and distribution line power defect image data into training data sets;
s202, using a training data set, and in an incremental training stage of a lifetime learning process, performing full-scale parameter updating training on a detection model with a characteristic pyramid network by using a lifetime learning model of model fine adjustment to obtain an electric power defect detection model.
3. The method for training and detecting a lifetime learning power model based on outlier rejection according to claim 2, wherein S202 comprises the steps of:
s2021, if the current training state is the initial training in the model fine-tuning lifetime learning paradigm process, entering S2022, otherwise, entering S2023;
s2022, extracting a back bone network by taking a ResNet50 network as a characteristic, extracting a Neck network by taking a characteristic pyramid network as a characteristic, and constructing a complete electric power defect detection model by combining a classification network and a bounding box regression network;
s2023, calling a power defect detection model used in the previous training in the lifetime learning process;
s2024, training the electric power defect detection model according to a lifetime learning paradigm of model fine adjustment to obtain the electric power defect detection model, wherein the training process comprises the following steps: activating network neurons in the power defect detection model by using a Relu activation function, performing maximum pooling operation on the power characteristic diagram extracted from the power defect detection model, and calculating loss on a classification network and a bounding box regression network in the power defect detection model.
4. The method for training and detecting a lifetime learning power model based on outlier rejection according to claim 3, wherein the expression for activating network neurons in the power defect detection model is as follows:
wherein,network neurons representing inputs, +.>Network neurons representing outputs, when +.>When the ratio is not more than 0,when->Greater than 0->，/>Representing a maximization operation.
5. The outlier rejection-based life-long learning power model training and detection method of claim 3, wherein the expression of the maximum pooling operation is as follows:
wherein,representation ofElectric defect detection model NojThe first power characteristic diagramkOutput value of individual pooling window, +.>Representing the first of the power defect detection modelsjThe first power characteristic diagramkThe first area covered by the pooling windowiThe value of the individual pixels>Representing the first of the power defect detection modelsjThe first power characteristic diagramkAreas covered by individual pooling windows, +.>Representing a maximization operation.
6. The method for training and detecting a lifetime learning power model based on outlier rejection according to claim 3, wherein the expressions for calculating the loss of classification network and bounding box regression network in the power defect detection model are as follows:
wherein,representing the loss function of the classification network,Nsample total number representing power defect image data, +.>Indicate->The true value of the individual power defect image data samples, is->Indicate->Predictive value of individual power defect image data samples, < >>Representing a logarithmic function>Respectively representing a set of predicted values and real values;
wherein,loss function representing a regression network of bounding boxes, +.>Representing the actual value +.>Representing the predicted value.
7. The method for training and detecting a lifetime learning power model based on outlier rejection according to claim 1, wherein S301 comprises the steps of:
s3011, acquiring parameter dimensionality of a feature pyramid module in the electric power defect detection modelWherein->Respectively representing the size of the parameter matrix in three dimensions of length, width and height;
s3012, traversing all parameters according to the parameter dimension by the feature pyramid module, and mapping all traversed parameters to a one-dimensional space through dimension mapping;
s3013, obtaining the original dimension information of the feature pyramid module according to the processing result of S3012.
8. The outlier rejection-based life-long learning power model training and detection method of claim 7, wherein the dimension map is calculated as:
wherein,representing dimension map->Respectively representing the coordinate values of the parameters in the dimension space.
9. The method for training and detecting an outlier rejection-based life-long learning power model according to claim 1, wherein S305 comprises the steps of:
s3051, set of lifetime parametersThe average division of the value range between the maximum and minimum values is +.>A plurality of regions;
s3052, collecting the lifetime parametersAll parameters in the power transmission and distribution line are distributed to corresponding parameter areas according to the values to obtain a plurality of power transmission and distribution line parameter areas +.>。
10. The outlier rejection-based life-long learning power model training and detection method of claim 9, wherein the partitioning isThe expression of the individual regions is as follows:
wherein,and->Respectively representing lifetime parameter sets->Maximum and minimum of (a), j->The number of divided regions is indicated,drepresenting the size of each region after division, +.>Indicate->Left end point of individual region,/>Indicate->The right end point of the region.
11. The outlier rejection-based lifelong learning power model training and detection method of claim 1, wherein the power fusion parametersThe expression of (2) is as follows:
wherein,nrepresenting power dense areasThe number of elements in the collection, +.>Representing a power dense region->Is a component of the group.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311526946.7A CN117274748B (en) | 2023-11-16 | 2023-11-16 | Lifelong learning power model training and detecting method based on outlier rejection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311526946.7A CN117274748B (en) | 2023-11-16 | 2023-11-16 | Lifelong learning power model training and detecting method based on outlier rejection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117274748A CN117274748A (en) | 2023-12-22 |
| CN117274748B true CN117274748B (en) | 2024-02-06 |
Family
ID=89208326
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311526946.7A Active CN117274748B (en) | 2023-11-16 | 2023-11-16 | Lifelong learning power model training and detecting method based on outlier rejection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117274748B (en) |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2012007952A (en) * | 2010-06-23 | 2012-01-12 | Panasonic Electric Works Co Ltd | Visual inspection device |
| GB201223057D0 (en) * | 2012-09-25 | 2013-02-06 | Lg Display Co Ltd | Display device and method for detecting line defects of the display device |
| CN105759177A (en) * | 2016-04-26 | 2016-07-13 | 浙江大学城市学院 | Classified-multi-mode-fusion-based distributed grid island detection method |
| CN109521330A (en) * | 2018-12-06 | 2019-03-26 | 中国矿业大学 | A kind of transmission line malfunction travelling wave ranging method based on the prediction of ARIMA wave head |
| CN114299354A (en) * | 2021-11-29 | 2022-04-08 | 上海电力大学 | Insulator detection positioning method based on rotating frame identification network |
| CN115018111A (en) * | 2022-01-24 | 2022-09-06 | 昆明理工大学 | Wind power prediction method and system integrating deep learning and self-adaptive modeling mechanisms |
| CN115272761A (en) * | 2022-07-26 | 2022-11-01 | 国网四川省电力公司电力科学研究院 | Power grid transmission line defect detection method based on collaborative representation bilateral self-encoder |
| CN115861263A (en) * | 2022-12-19 | 2023-03-28 | 安徽南瑞继远电网技术有限公司 | Insulator defect image detection method based on improved YOLOv5 network |
| CN116228754A (en) * | 2023-05-08 | 2023-06-06 | 山东锋士信息技术有限公司 | Surface defect detection method based on deep learning and global difference information |
| CN116468666A (en) * | 2023-03-14 | 2023-07-21 | 贵州电网有限责任公司 | Inspection image defect detection method special for operation and maintenance of power transmission line |
| CN116486305A (en) * | 2023-04-27 | 2023-07-25 | 中国海洋大学 | Image recognition-based deep sea suspended particulate matter concentration prediction method |
| CN116843650A (en) * | 2023-07-04 | 2023-10-03 | 上海交通大学 | SMT welding defect detection method and system integrating AOI detection and deep learning |
| CN116994047A (en) * | 2023-08-01 | 2023-11-03 | 北京工商大学 | Small sample image defect target detection method based on self-supervision pre-training |
-
2023
- 2023-11-16 CN CN202311526946.7A patent/CN117274748B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2012007952A (en) * | 2010-06-23 | 2012-01-12 | Panasonic Electric Works Co Ltd | Visual inspection device |
| GB201223057D0 (en) * | 2012-09-25 | 2013-02-06 | Lg Display Co Ltd | Display device and method for detecting line defects of the display device |
| CN105759177A (en) * | 2016-04-26 | 2016-07-13 | 浙江大学城市学院 | Classified-multi-mode-fusion-based distributed grid island detection method |
| CN109521330A (en) * | 2018-12-06 | 2019-03-26 | 中国矿业大学 | A kind of transmission line malfunction travelling wave ranging method based on the prediction of ARIMA wave head |
| CN114299354A (en) * | 2021-11-29 | 2022-04-08 | 上海电力大学 | Insulator detection positioning method based on rotating frame identification network |
| CN115018111A (en) * | 2022-01-24 | 2022-09-06 | 昆明理工大学 | Wind power prediction method and system integrating deep learning and self-adaptive modeling mechanisms |
| CN115272761A (en) * | 2022-07-26 | 2022-11-01 | 国网四川省电力公司电力科学研究院 | Power grid transmission line defect detection method based on collaborative representation bilateral self-encoder |
| CN115861263A (en) * | 2022-12-19 | 2023-03-28 | 安徽南瑞继远电网技术有限公司 | Insulator defect image detection method based on improved YOLOv5 network |
| CN116468666A (en) * | 2023-03-14 | 2023-07-21 | 贵州电网有限责任公司 | Inspection image defect detection method special for operation and maintenance of power transmission line |
| CN116486305A (en) * | 2023-04-27 | 2023-07-25 | 中国海洋大学 | Image recognition-based deep sea suspended particulate matter concentration prediction method |
| CN116228754A (en) * | 2023-05-08 | 2023-06-06 | 山东锋士信息技术有限公司 | Surface defect detection method based on deep learning and global difference information |
| CN116843650A (en) * | 2023-07-04 | 2023-10-03 | 上海交通大学 | SMT welding defect detection method and system integrating AOI detection and deep learning |
| CN116994047A (en) * | 2023-08-01 | 2023-11-03 | 北京工商大学 | Small sample image defect target detection method based on self-supervision pre-training |
Non-Patent Citations (6)
| Title |
|---|
| Multi-Stage Feature Enhancement Pyramid Network for Detecting Objects in Optical Remote Sensing Images;Kaihua Zhang等;《Remote Sensing》;第14卷(第3期);第1-22页 * |
| Transmission Line Fault Detection and Classification Based on Improved YOLOv8s;Hao Qiang等;《Electronics》;第12卷(第21期);第1-18页 * |
| 基于可变形卷积神经网络的遥感影像密集区域车辆检测方法;高鑫 等;《电子与信息学报》;第40卷(第12期);第2812-2819页 * |
| 基于深度学习的液晶屏Mura缺陷图像识别算法研究;梅爽;《中国博士学位论文全文数据库 信息科技辑》(第1期);第I136-28页 * |
| 基于深度学习的输电线路故障类型辨识;徐舒玮 等;《中国电机工程学报》;第39卷(第1期);第65-74页 * |
| 融合结构推理的深度模型输电线路金具及其缺陷检测;赵振兵 等;《高电压技术》;第49卷(第8期);第3346-3353页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117274748A (en) | 2023-12-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Corchado et al. | Data mining using example-based methods in oceanographic forecast models | |
| CN107967542B (en) | Long-short term memory network-based electricity sales amount prediction method | |
| CN110648014B (en) | Regional wind power prediction method and system based on space-time quantile regression | |
| CN111814570B (en) | Face recognition method, system and storage medium based on dynamic threshold | |
| CN109615124B (en) | SCADA master station load prediction method based on deep learning | |
| CN114898121B (en) | Automatic generation method for concrete dam defect image description based on graph attention network | |
| CN110599459A (en) | Underground pipe network risk assessment cloud system based on deep learning | |
| CN113361801A (en) | Hybrid rolling short-term load prediction method based on GAN network and semi-supervised regression | |
| KR20220078963A (en) | Deep learning based photovoltaic power generation forecasting method using satellite images and apparatus using the same | |
| CN110599458A (en) | Underground pipe network detection and evaluation cloud system based on convolutional neural network | |
| CN114037140A (en) | Prediction model training method, prediction model training device, prediction model data prediction method, prediction model data prediction device, prediction model data prediction equipment and storage medium | |
| CN113205223A (en) | Electric quantity prediction system and prediction method thereof | |
| CN113361803A (en) | Ultra-short-term photovoltaic power prediction method based on generation countermeasure network | |
| CN117274748B (en) | Lifelong learning power model training and detecting method based on outlier rejection | |
| CN113177355B (en) | Power load prediction method | |
| CN101414365A (en) | Vector code quantizer based on particle group | |
| CN117155806A (en) | Communication base station flow prediction method and device | |
| CN116596169A (en) | Power system prediction method, device and storage medium | |
| CN117131654A (en) | Target observation method based on nonlinear optimal disturbance of pre-analysis initial guess condition | |
| CN115660326A (en) | Power system standby management method, device, storage medium and system | |
| CN109034453A (en) | A kind of Short-Term Load Forecasting Method based on multiple labeling neural network | |
| CN115907131A (en) | Method and system for building electric heating load prediction model in northern area | |
| CN115170528A (en) | Pavement defect detection method, system, equipment and storage medium | |
| CN112380985A (en) | Real-time detection method for intrusion foreign matters in transformer substation | |
| CN110751335B (en) | Regional ecological quality annual scene prediction evaluation method and device |
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 |