CN117994630A - Consistency teacher learning model applied to transmission line anti-external damage monitoring system - Google Patents
Consistency teacher learning model applied to transmission line anti-external damage monitoring system Download PDFInfo
- Publication number
- CN117994630A CN117994630A CN202410093965.3A CN202410093965A CN117994630A CN 117994630 A CN117994630 A CN 117994630A CN 202410093965 A CN202410093965 A CN 202410093965A CN 117994630 A CN117994630 A CN 117994630A
- Authority
- CN
- China
- Prior art keywords
- model
- teacher
- inconsistent
- data set
- ssod
- 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
- 238000012544 monitoring process Methods 0.000 title claims abstract description 14
- 230000005540 biological transmission Effects 0.000 title claims abstract description 13
- 238000011156 evaluation Methods 0.000 claims abstract description 16
- 238000000034 method Methods 0.000 claims abstract description 10
- 238000002474 experimental method Methods 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 13
- 238000005457 optimization Methods 0.000 claims description 10
- 239000000203 mixture Substances 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 5
- 125000004122 cyclic group Chemical group 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 2
- 238000009826 distribution Methods 0.000 abstract description 11
- 230000000750 progressive effect Effects 0.000 abstract 1
- 238000001514 detection method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a consistency teacher learning model applied to an anti-external damage monitoring system of a power transmission line, which selects MSCOCO and PASCAL VOC data sets as samples to carry out comprehensive experiments and sets an evaluation scheme, takes a general SSOD model as a base line, enables the teacher model to be kept as an index moving average value of a student detector, sequentially inputs unlabeled parts of the data sets into the teacher model through weak enhancement to generate pseudo marks, supervises a student network by the pseudo marks, enables the same unlabeled samples to shake strongly, simultaneously inputs a small part of labeled data into the student network to improve the performance of the student network, sets the goal of the student network to minimize loss, optimizes the problems of inconsistent distribution, inconsistent subtask and inconsistent time in the SSOD model, trains the data sets by the model, and can obtain the model with the performance superior to that of the current progressive method.
Description
Technical Field
The invention relates to the technical field of machine vision, in particular to a consistent teacher learning model applied to an anti-external damage monitoring system of a power transmission line.
Background
With the development of science and technology and national economy, the electric power consumption is increased, the laying amount of the electric transmission lines in China is gradually increased, and accidents caused by the damage of the electric transmission lines by external force are more and more. In order to solve the accident, the manual early warning consumes more manpower and material resources, has high false alarm rate of missing report and can not accurately and timely transmit accident information; the existing automatic monitoring alarm equipment has high false alarm rate, is easily interfered by external factors, and cannot accurately convey accident information. In recent years, the domestic computer vision technology has been gradually applied to the electrical industry, and electrical equipment can be monitored in real time by using a vision system. Although the electrical equipment can acquire a large number of sample graphs for learning, the equipment has the defects of long learning time, high learning cost, large influence by factors such as environment and the like, easy occurrence of problems of sinking into local optimal solution and low recognition accuracy in monitoring. The semi-supervised algorithm for equipment vision learning has the advantages of less learning cost and higher accuracy, but has the problems of inconsistent allocation, inconsistent subtasks and inconsistent time.
Disclosure of Invention
The invention mainly solves the technical problems existing in the prior art; the device for detecting the external damage is trained by the consistent teacher learning model, so that the learning cost of the device can be greatly reduced, and the monitoring accuracy of the device is improved.
Still another object of the present invention is to solve the technical problems of the prior art; the consistency teacher learning model can optimize the existing semi-supervised learning algorithm, and solves the problems of inconsistent distribution, inconsistent subtasks and inconsistent time in the semi-supervised learning algorithm.
The technical problems of the invention are mainly solved by the following technical proposal:
the consistent teacher learning model applied to the transmission line anti-external damage monitoring system is characterized by comprising the following steps of:
Step1, a data set and evaluation setting step: for verifying performance indexes of consistency teachers, MSCOCO and PASCAL VOC datasets are selected as samples to carry out comprehensive experiments, and three evaluation schemes are set at the same time;
Step 2, SSOD frame construction: the universal SSOD paradigm is used as a base line, a teacher model is kept as an index moving average value of a student detector, the unlabeled part of the dataset in the step 1 is input into the teacher model through weak enhancement in sequence to generate a pseudo-label, the student network is supervised by the pseudo-label, the same unlabeled sample is subjected to strong jitter, at the same time, a small part of labeled data is input into the student network to improve the performance of the student network, and the goal of the student network is set to minimize loss;
Step 3, SSOD inconsistent problem optimization step: for the problem of inconsistent allocation, self-adaptive sample allocation is adopted to replace anchor point allocation based on the IOU, the matching cost between each predicted value and a given true value is calculated, and allocation is carried out according to the matching cost, so that the optimized consistency is enhanced; for the problem of inconsistent subtasks, a 3D feature symmetry module is referenced, and different feature positions are adaptively selected to complete each subtask so as to estimate the offset of the optimal regression feature points and enable the optimal regression feature points to be better aligned with the classification branches; for the problem of inconsistent time, a Gaussian mixture model is adopted to adaptively adjust a threshold value, and the strategy reduces the complicated super-parameter adjustment requirement in SSOD, so that the robust improvement is realized under different data sets and settings;
step 4, model training step: dividing the data set in the step 1 into a small data set with a mark and a large data set without a mark, inputting the small data set with the mark into the optimization model obtained in the step 3 for preliminary training, inputting the large data set without the mark, adding the sample with higher output confidence to the data set with the mark, and performing cyclic training by the method;
step 5, model test step: under the three evaluation modes described in the step 1, the performance of the optimization model obtained in the step 3 is calculated and compared with the performance of the current most advanced method.
In the foregoing consistent teacher learning model applied to the transmission line anti-external damage monitoring system, in the step 1, three evaluation schemes are set, and the evaluation schemes are respectively as follows: COCO-PARTIAL, COCO-ADDITION and VOC-PARTIAL.
In the above-mentioned consistent teacher learning model applied to the transmission line anti-external damage monitoring system, in the step 5, the performance of the model is compared with that of the current most advanced method, which is a Mean-Teacher model with RETINANET detector.
The consistency teacher learning model is characterized by comprising an optimization method for solving SSOD inconsistent problems, wherein the self-adaptive sample distribution is used for solving the inconsistent problems of distribution, the 3D feature symmetry module is used for solving the inconsistent problems of subtasks, and the Gaussian mixture model is used for solving the inconsistent problems of time.
Therefore, the invention has the following advantages: 1. the SSOD algorithm is adopted to train the transmission line anti-external-damage monitoring system, so that the learning cost is greatly reduced, and the anti-interference capability and the environmental adaptability of the system are improved; 2. the problem of inconsistency in the traditional SSOD is solved by adopting the self-adaptive sample distribution, the 3D characteristic symmetry module and the Gaussian mixture model, and the performance index of the algorithm is improved.
Drawings
FIG. 1 is a workflow diagram of a model of the present invention.
FIG. 2 is a schematic view of a model test head according to the present invention, wherein the left side is a head structure and the right side is a 3-DFAM module.
Fig. 3 is a score distribution and fitting result of two gaussian mixtures. The positive and negative distributions are indicated by orange and blue lines, respectively, and the blue dashed line indicates the final threshold.
Detailed Description
The technical scheme of the invention is further specifically described below through examples and with reference to the accompanying drawings.
Examples:
A consistency teacher learning model applied to an anti-outward-breakage monitoring system of a power transmission line comprises the following steps:
step 1, in order to verify performance indexes of consistency teachers, MSCOCO and PASCAL VOC datasets are selected as samples to carry out comprehensive experiments, and three evaluation schemes are set at the same time, wherein the evaluation schemes are respectively as follows: to ensure fair comparison, all detectors were trained on 8 GPUs, 5 images per GPU (1 marked image and 4 unmarked images), detector optimized with SGD, constant learning rate 0.01, momentum 0.9, weight decay 0.001, unmarked data weight λ u =2, teacher model updated by EMA, momentum 0.999, RETINANET set as default detection framework, and ResNet-50-FPN as its backbone, imageNet pre-training model as initialization, the specific method of operation for the three evaluation schemes is as follows:
Step 1.1, for COCO-PARTIAL, randomly sampling 1%/2%/5%/10% of the image in the train2017 as the marker data, and taking the result of the AP 50:95 in the VAL2017 as the evaluation index;
Step 1.2, for COCO-ADDITION, using the complete train2017 as the labeled set, using the COCO official unlabeled set unlabel2017 as the unlabeled set, training the model;
Step 1.3, for VOC-partal, use the VOC2007 trainval set as labeled data, and use VOC2012 trainval as unlabeled data. The final model was validated on the VOC2007 test set using AP 50 and AP 50:95;
Step 2, using a general SSOD paradigm as a base line, keeping a teacher model as an index moving average value of a student detector, sequentially inputting an unlabeled part of the data set in step 1 into the teacher model through weak enhancement to generate a pseudo-label, supervising a student network by the pseudo-label, strongly dithering the same unlabeled sample, simultaneously inputting a small part of the labeled data into the student network to improve the performance of the student network, setting a target of the student network as a minimum loss, wherein the loss formula of the student network is as follows:
Where L s and L u represent supervised and unsupervised (unlabeled) losses, respectively, lambda u is the loss weight of L u and lambda reg is the loss weight of the regression subtask.
Step 3, for the problem of inconsistent allocation, adopting self-adaptive sample allocation to replace anchor point allocation based on the IOU, calculating the matching cost between each predicted value and a given true value, and allocating according to the matching cost, thereby enhancing the optimized consistency; for the problem of inconsistent subtasks, a 3D feature symmetry module is referenced, and different feature positions are adaptively selected to complete each subtask so as to estimate the offset of the optimal regression feature points and enable the optimal regression feature points to be better aligned with the classification branches; for the problem of inconsistent time, a Gaussian mixture model is adopted to adaptively adjust a threshold value, and the strategy reduces the complicated super-parameter adjustment requirement in SSOD, so that the robust improvement is realized under different data sets and settings; the specific operation method of the three optimization schemes is as follows:
Step 3.1, for the problem of inconsistent allocation, an adaptive sample allocation strategy is adopted, a feature point p i is given, a predicted value pred bbox is obtained by a detector, and the cost between each gt b i and the prediction of the feature point p j is calculated, wherein the cost formula is as follows:
Cij＝λclsCcls+λregCreg+λdistCdist
wherein C cls＝Lcls(Pred(pi)cls,bi)
Creg＝Lreg(Pred(pi)reg,bi)
Where C cls and C reg represent classification and regression costs, respectively, C dist (i, j) represents center priors, i d (p j,bi) i represents the distance of the feature point p j to the center of b i, and λ cls、λreg and λ dist are weighting parameters by step-wise normalization in the FPN level.
After the matching costs between b i and all the feature points p j are sorted out, the first K feature points with lost costs are regarded as positive candidates, and since C dist (i, j) is mainly stable distribution and lambda dist is smaller, the matching cost is still dominated by classification and regression quality, and since distribution is carried out according to detection quality, the influence of noise in a pseudo frame on the feature point distribution is negligible;
Step 3.2, for the subtask inconsistency problem, referring to a 3D feature symmetry module, predicting the decoded bbox coordinates offset in the 3D dimension by an offset prediction module, adding an additional branch, the input of which is the cascade middle layer of cls and reg branches, the mapping of the cascade feature being f e R H×W×8C, wherein c=256 represents the number of channels of the middle feature, reducing the dimension to f ' er H×W×C/4 by convolution of 1X1, retaining f ' in each FPN level to predict the 3D offset, then interpolating f ' l-1 and f ' l+1 to the same size as f l ', wherein l refers to the index of the FPN level, connecting the features of different FPN levels to output offset predictions O e R H×W×12, each boundary prediction feature point P (i; j; l) being able to find the respective offset, wherein i, j is the position index, the offset is in 3D space, expressed as d= (D 1,d2,d3), wherein D 1(i,j,l)＝O(i,j,3c+1),d2(i,j,l)＝O(i,j,3c+2),d3 (i, j) =3C, j is the two consecutive steps of the offset:
Ph(i,j,l)←P(i+d1,j+d2,l)
Step 3.3, for the problem of inconsistent time, a Gaussian mixture model is adopted, a probability rule is used for distinguishing a positive pseudo frame from a negative pseudo frame, and a confidence score is obtained from the mixture of Gaussian distribution, wherein the confidence score formula is as follows:
P(b|s,θ)＝wnNn(b,μn,pn)+wpNp(b,μp,pp)
where b represents confidence scores of the artifacts bbox, w n、μn、pn and w p、μp、pp represent weights, means and accuracy of two gaussian models, respectively, a confidence score library of size N is built for each model, the GMM is optimized in each step using a expectation maximization algorithm, the probability of each artifact bbox being positive or negative is determined by the parameters of the GMM, the threshold τ is determined by the sample with the highest positive probability, and the formula of τ is:
By fitting the GMM using the latest N samples, the threshold can be adaptively adjusted;
Step 4, dividing the data set in the step 1 into a small data set with a mark and a large data set without a mark, inputting the small data set with the mark into the optimization model obtained in the step 3 for preliminary training, inputting the large data set without the mark, adding the sample with higher output confidence into the data set with the mark, and performing cyclic training through the method;
step 5, model test step: under the three evaluation modes described in the step 1, the performance of the optimization model obtained in the step 3 is calculated, and compared with a Mean-Teacher model with RETINANET detectors, and the comparison results of the three optimization schemes are as follows:
Step 5.1, for the COCO-PARTIAL comparison result, after training the COCO 1%/2%/5%/10% label, comparing the AP 50:95 of the two models, it can be seen that the Mean-Teacher model with RETINANET detector has an AP 50:95 value of 35.5 on the COCO 10% experiment, which shows the powerful performance of SSOD algorithm, and the Constant-Teacher obtains an AP 50:95 value of 40.0, which is more than the Mean-Teacher model with RETINANET detector;
Step 5.2, comparing the Constant-Teacher model with other SSOD methods on the VOC0712 dataset for the VOC-partal comparison result, note that significant improvement was made to the Constant-Teacher model, which increased the mAP by 7.9.
And 5.3, regarding a COCO-ADDITION comparison result, taking the complete train2017 as marked data, taking the extra unlabel2017 as unmarked data, taking ResNet-50 as a main body of a Constant-Teacher model, and carrying out 360K iterative training, wherein the AP 50:95 value reaches 49.10, and compared with PseCo, namely a current most advanced semi-supervised detector, the Constant-Teacher model can reach the AP 50:95 value of 3.0 only by half training time.
Claims (2)
1. The consistent teacher learning model applied to the transmission line anti-external damage monitoring system is characterized by comprising the following steps of:
Step1, a data set and evaluation setting step: for verifying performance indexes of consistency teachers, MSCOCO and PASCAL VOC datasets are selected as samples to carry out comprehensive experiments, and three evaluation schemes are set at the same time;
Step 2, SSOD frame construction: the universal SSOD paradigm is used as a base line, a teacher model is kept as an index moving average value of a student detector, the unlabeled part of the dataset in the step 1 is input into the teacher model through weak enhancement in sequence to generate a pseudo-label, the student network is supervised by the pseudo-label, the same unlabeled sample is subjected to strong jitter, at the same time, a small part of labeled data is input into the student network to improve the performance of the student network, and the goal of the student network is set to minimize loss;
Step 3, SSOD inconsistent problem optimization step: for the problem of inconsistent allocation, self-adaptive sample allocation is adopted to replace anchor point allocation based on the IOU, the matching cost between each predicted value and a given true value is calculated, and allocation is carried out according to the matching cost, so that the optimized consistency is enhanced; for the problem of inconsistent subtasks, a 3D feature symmetry module is referenced, and different feature positions are adaptively selected to complete each subtask so as to estimate the offset of the optimal regression feature points and enable the optimal regression feature points to be better aligned with the classification branches; for the problem of inconsistent time, a Gaussian mixture model is adopted to adaptively adjust a threshold value, and the strategy reduces the complicated super-parameter adjustment requirement in SSOD, so that the robust improvement is realized under different data sets and settings;
step 4, model training step: dividing the data set in the step 1 into a small data set with a mark and a large data set without a mark, inputting the small data set with the mark into the optimization model obtained in the step 3 for preliminary training, inputting the large data set without the mark, adding the sample with higher output confidence to the data set with the mark, and performing cyclic training by the method;
step 5, model test step: under the three evaluation modes described in the step 1, the performance of the optimized model obtained in the step 3 is calculated and compared with the performance of the Mean-Teacher model with RETINANET detectors.
2. The consistency teacher learning model applied to the transmission line anti-external damage monitoring system according to claim 1, wherein in the step 1, three evaluation schemes are set, and the evaluation schemes are respectively: COCO-PARTIAL, COCO-ADDITION and VOC-PARTIAL.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410093965.3A CN117994630A (en) | 2024-01-23 | 2024-01-23 | Consistency teacher learning model applied to transmission line anti-external damage monitoring system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410093965.3A CN117994630A (en) | 2024-01-23 | 2024-01-23 | Consistency teacher learning model applied to transmission line anti-external damage monitoring system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117994630A true CN117994630A (en) | 2024-05-07 |
Family
ID=90886581
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410093965.3A Pending CN117994630A (en) | 2024-01-23 | 2024-01-23 | Consistency teacher learning model applied to transmission line anti-external damage monitoring system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117994630A (en) |
-
2024
- 2024-01-23 CN CN202410093965.3A patent/CN117994630A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598029B (en) | Fine-grained image classification method based on attention transfer mechanism | |
| CN112150821B (en) | Lightweight vehicle detection model construction method, system and device | |
| CN109117883B (en) | SAR image sea ice classification method and system based on long-time memory network | |
| WO2018150812A1 (en) | Balancing active learning | |
| CN114299380A (en) | Remote sensing image semantic segmentation model training method and device for contrast consistency learning | |
| CN111860106B (en) | Unsupervised bridge crack identification method | |
| CN109581339B (en) | Sonar identification method based on automatic adjustment self-coding network of brainstorming storm | |
| CN111209832B (en) | Auxiliary obstacle avoidance training method, equipment and medium for substation inspection robot | |
| CN112766218B (en) | Cross-domain pedestrian re-recognition method and device based on asymmetric combined teaching network | |
| CN113139594A (en) | Airborne image unmanned aerial vehicle target self-adaptive detection method | |
| CN111582358B (en) | Training method and device for house type recognition model, and house type weight judging method and device | |
| CN113642486A (en) | Unmanned aerial vehicle distribution network inspection method with airborne front-end identification model | |
| CN112633149A (en) | Domain-adaptive foggy-day image target detection method and device | |
| CN112580575A (en) | Electric power inspection insulator image identification method | |
| CN115471712A (en) | Learning method for generating zero sample based on visual semantic constraint | |
| CN111144462A (en) | Unknown individual identification method and device for radar signals | |
| CN110751005B (en) | Pedestrian detection method integrating depth perception features and kernel extreme learning machine | |
| CN114549909A (en) | Pseudo label remote sensing image scene classification method based on self-adaptive threshold | |
| CN117274212A (en) | Bridge underwater structure crack detection method | |
| CN116977710A (en) | Remote sensing image long tail distribution target semi-supervised detection method | |
| CN117994630A (en) | Consistency teacher learning model applied to transmission line anti-external damage monitoring system | |
| CN115797904A (en) | Active learning method for multiple scenes and multiple tasks in intelligent driving visual perception | |
| CN115331135A (en) | Method for detecting Deepfake video based on multi-domain characteristic region standard score difference | |
| CN114283336A (en) | Anchor-frame-free remote sensing image small target detection method based on mixed attention | |
| CN110728292A (en) | Self-adaptive feature selection algorithm under multi-task joint optimization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination |