Research On Separation And Identification Of Air Leakage Signal Based On Deep Learning

The leakage of harmful gases has always been a problem worthy of attention,such as carbon dioxide chemical products sealed in the engine room of ships,liquid ammonia liquid nitrogen,natural gas,sulfur hexafluoride arc extinguishing gas,etc.If the leakage occurs,it will not only cause pollution to the atmosphere,but also may pose a threat to the life safety of staff.In order to improve the effectiveness of the detection of harmful gas leakage,the paper takes the sound signal of gas leakage as the object,collects the leakage signal,develops the optimized Wave-U-Net deep learning network for noise reduction,converts the sound signal into idiom spectrum,studies the optimized Res Net deep neural network recognition spectrum,designs the combined network.The system solution of separation and identification of gas leakage signal is given,and the identification of gas leakage is verified by taking sulfur hexafluoride gas in GIS equipment as an example.The main work of this paper includes:(1)Study the identification method of air leakage signal based on improved Res Net.Firstly,the applicability of the network model to sound source signals of different acquisition frequencies is studied,and the training set for air leakage identification required by network training is developed: Aiming at the problem that one-dimensional acoustic signals are input to convolutional neural network as images,the acoustic signals collected by pickups with different sampling rates are preprocessed by resampling.After the frame length and slider length are set,the fast Fourier transform of audio files is converted to time-frequency spectrum to connect with the input interface of Res Net network.Then,CBAM attention module is added to Res Net network to optimize the network.The optimized network pays more attention to the key part of the feature graph to be extracted,which effectively improves the network performance.At the same time,the output layer of the network is optimized according to the task requirements of this paper,and the learning rate,the number of training per batch and other super parameters are optimized by means of comparative training.Finally,the effectiveness of the model in identifying air leakage signals is proved by field experiments.(2)Study the separation method of leakage signal based on improved Wave-Unet.Considering the shortage of the model of weak leakage signal recognition,data set is first proposed to improve the structure of the scheme,the pure air leakage through gathering signal and the possible interference in actual scene background noise signal,the attenuation factor combined with adaptive audio mixing method and the background noise sound power of gain,A data set was developed to solve the problem of weak signal recognition.Then,a method of suppressing complex substation background noise interference is proposed in the leakage signal identification network in front of the leakage signal separation network.Wave-U-Net network is used as the pre-acoustic source separation and noise reduction network,and a training network model of training data set is developed to increase the noise gain of different background interference.A new loss function is proposed combining the evaluation index of blind source separation and local loss.Finally,the experimental results show that the network performance has been improved.(3)Research on noise reduction recognition method of air leakage signal based on improved Res Net & Wave-U-Net combined network.Put forward by a front Wave-UNet noise purification network and rear Res Net leak signal to identify the combination of network to do leak signal separation and identification of system solutions,through the experiment of the substation site prove model easy to identify to the human ears loud air leakage signal in the original good enough recognition on the basis of ascension and to a certain extent,For the weak leakage signal which is difficult to be distinguished by human ears,the original unsatisfactory recognition rate has been significantly improved.The substation sulfur hexafluoride leakage detection problem and deep learning method can be combined to detect whether the leakage occurs in real time.