Recently, people’s demand for action recognition has extended from the initial high classification accuracy to the high accuracy of the temporal action detection. It is challenging to meet the two requirements simultaneously. The key to behavior recognition lies in the quantity and quality of the extracted features. In this paper, a two-stream convolutional network is used. A three-dimensional convolutional neural network (3D-CNN) is used to extract spatiotemporal features from the consecutive frames. A two-dimensional convolutional neural network (2D-CNN) is used to extract spatial features from the key-frames. The integration of the two networks is excellent for improving the model’s accuracy and can complete the task of distinguishing the start–stop frame. In this paper, a multi-scale feature extraction method is presented to extract more abundant feature information. At the same time, a multi-task learning model is introduced. It can further improve the accuracy of classification via sharing the data between multiple tasks. The experimental result shows that the accuracy of the modified model is improved by 10%. Meanwhile, we propose the confidence gradient, which can optimize the distinguishing method of the start–stop frame to improve the temporal action detection accuracy. The experimental result shows that the accuracy has been enhanced by 11%.

these two kinds of detectors are still in its infancy and have Abstract-Electrochemical sensors and gas chmmatography ' encountered many problems that optical detecting methods are widely used in on-line monitors for dissolved gas-in-oil of power transformers, but some fatal defects exist. In view of the many advantages of FTIR, the basics of applying it on such monitors are studied and results are prcsented. Investigation shows that ail the characteristic Deal\s needed to be analwed could avoid naturally. The most outstanding advantages of an optical detector are their non.destruct,ve and no characteristics; and with FTIR spectroscopy, it is reported'"' that the amount of all the diagnostic gases as well as water can be covered if the measurable wavenumber rcgion of the spectrometer is 720cm.'-3400cm.', and the optimum optical content except hydrogen could be measured slmultanwusly m a satisfactory precision that such B monitor'can. be consiructed using FTIR that can FTIR spectroscopy are investigated recently'"', no rePo* on Dreciselv detect all the diamostic eases excent hvdrocen and are FTIR suectra used in DGA has been made in China. " 0 . ~ ~. free of consuming gases as well as in situ calibration.Index Term-DGA, dissolved gas-in-oil, FTIR, gas cell, infrared spectra, on-line monitoring, optical length, power transformer.

No abstract

Recently, chromatography column and gas sensor have been used in online monitoring device of dissolved gases in transformer oil. But some disadvantages still exist in these devices: consumption of carrier gas, requirement of calibration, etc. Since FTIR has high accuracy, consume no carrier gas and require no calibration, the researcher studied the application of FTIR in such monitoring device. Experiments of "Flow gas method" were designed, and spectrum of mixture composed of different gases was collected with A BOMEM MB104 FTIR Spectrometer. A key question in the application of FTIR is that: the absorbance spectrum of 3 fault key gases, including C 2 H 4 , CH 4 and C 2 H 6 , are overlapped seriously at 2700~3400cm -1 . Because Absorbance Law is no longer appropriate, a nonlinear calibration model based on BP ANN was setup to in the quantitative analysis. The height absorbance of C 2 H 4 , CH 4 and C 2 H 6 were adopted as quantitative feature, and all the data were nomalized before training the ANN. Computing results show that the calibration model can effectively eliminate the cross disturbance to measurement.

Recently, Semiconductor sensor and thermal conductivity sensor are widely used for gas detection in transformer online monitors. Since the long-time stability or precision of these sensors is not satisfactory, the present researcher studied the application of FTIR in such monitors. In the wide measuring range of online monitoring, Absorbance Law is not always applicable, thus a non-linear calibration model was necessary. Experiments were done to set up the calibration model. A gas dilution system was designed. With the system, standard samples of fault gas including CH 4 , C 2 H 2 , C 2 H 4 and C 2 H 6 were diluted to different concentration. BOMEM MB104 FTIR Spectrometer was used to collect spectra of gases. Curve fitting of the output of FTIR was done, and the effect of quantitative feature and concentration range on quantitative analysis was investigated. In addition, the lowest detection limit was tested. Experiment and calculation results show: accuracy can be improved by taking strong peak height at low concentration range, taking peak area or weak peak height at high concentration range as quantitative feature, and using third order polynomial to fit the output curve of FTIR. The lowest detecting limit of C 2 H 2 with 2.4m gas cell is below 0.3µl/l and that of 10cm cell is below 3µl/l.