阅读说明：本技术 一种基于Mel语图与深度学习的电网涉鸟故障鸟种识别方法 (Electric network bird-involved fault bird species identification method based on Mel language graph and deep learning ) 是由 邱志斌 卢祖文 廖才波 王海祥 于 2021-08-02 设计创作，主要内容包括：本发明公开了一种基于Mel语图与深度学习的电网涉鸟故障鸟种识别方法。首先建立电网涉鸟故障相关鸟种的鸣声样本数据库,对鸟鸣信号进行预处理操作后,计算每帧信号在各个Mel滤波器中的能量,得到包含信号能量大小信息的M×N矩阵,将能量大小与颜色深浅程度一一映射,得到鸟鸣信号的Mel语图。通过Mel语图训练卷积神经网络,执行卷积-池化过程不断地抓取学习鸟鸣信号的Mel语图特征,通过多次迭代训练调整网络内部参数,当网络的预测输出值与实际值之间的损失达到最小时结束训练,最终实现对测试鸟种的预测识别。该方法能够有效区分不同鸟种鸣声之间的特征并实现鸟种识别,可为开展电网涉鸟故障差异化防治提供参考。(The invention discloses a method for identifying bird species with bird-involved faults in a power grid based on Mel language maps and deep learning. Firstly, establishing a singing sample database of bird species related to bird-related faults of a power grid, preprocessing the singing signals, calculating the energy of each frame of signal in each Mel filter to obtain an M multiplied by N matrix containing signal energy information, and mapping the energy and the color depth one by one to obtain a Mel language map of the singing signals. The convolutional neural network is trained through the Mel tone map, the Mel tone map characteristic of the bird singing signal is continuously captured and learned in the convolution-pooling process, the internal parameters of the network are adjusted through repeated iterative training, the training is finished when the loss between the predicted output value and the actual value of the network is minimum, and finally the prediction and identification of the tested bird species are realized. The method can effectively distinguish the characteristics of different bird song sounds, realize bird identification and provide reference for carrying out bird-related fault differential control on the power grid.)

1. A method for identifying bird species with bird-involved faults in a power grid based on Mel language maps and deep learning is characterized by comprising the following steps: the method comprises the following steps:

s1: establishing a related bird seed singing database according to the counted main bird seeds with bird-involved faults of the power grid and the actual situation of the power grid;

s2: carrying out denoising, framing and windowing preprocessing on samples in a singing database, removing noise in the singing signal by adopting an improved spectral subtraction method of multi-window spectral estimation, framing the singing signal by setting the frame length and the frame shift, and multiplying the frame length and the frame shift by a window function to increase the continuity of two ends of the frame;

s3: calculating the energy of each frame of the birdsong signal in each Mel filter to obtain the Mel energy of the birdsong sample, obtaining an MxN order matrix containing signal energy information, mapping the energy and the color depth one by one to obtain a Mel language map of the birdsong signal, and dividing the Mel language map into a training set, a verification set and a test set;

s4: building a convolutional neural network classification model, performing repeated iterative training by taking a Mel (Mel) morphogram of a training set as input, testing a verification set in the training process to adjust parameters of the model, and finishing the training when the loss between a predicted output value and an actual value of the network reaches the minimum;

s5: and predicting and identifying the bird species in the test set by using the trained network, and outputting the corresponding bird species category.

2. The method for identifying the bird species with bird-involved fault in the power grid based on the Mel language map and the deep learning as claimed in claim 1, wherein: the calculation process of the Mel language graph in S3 is as follows: for a section of M-frame birdsong signals, N Mel filters are arranged, an MxN matrix is obtained through Mel energy calculation, a Mel language graph is obtained through coloring according to the energy, horizontal and vertical coordinates in the Mel language graph are the number of frames and the number of the filters respectively, only the MxN data size needs to be calculated, output is simplified, and meanwhile calculation time is shortened.

3. The method for identifying the bird species with bird-involved fault in the power grid based on the Mel language map and the deep learning as claimed in claim 1, wherein: and S4, the convolutional neural network comprises a plurality of convolution-pooling processes for capturing Mel language map features, network parameters and network iteration times are adjusted to train a training set, when the model is trained for a certain turn, the model carries out prediction on a verification set once and correspondingly adjusts the parameters according to the prediction result of the verification set, and the parameters are corrected in the direction with high prediction accuracy until the loss function value of the network is reduced to the minimum, and the network training is finished.

Technical Field

The invention relates to the field of power transmission lines, in particular to a method for identifying bird species with bird-involved faults in a power grid based on Mel language maps and deep learning.

Background

Bird activity is one of the important causes of overhead transmission line faults, although various bird prevention devices are widely applied, the bird prevention devices still have great blindness, the rising trend of bird-related faults cannot be effectively inhibited, and line tripping faults caused by the failure of the bird prevention devices also occur frequently. In addition, because bird-involved faults are transient, operation and maintenance personnel often have difficulty in judging the bird species causing the faults after the faults occur, an intelligent bird species identification and fault cause judgment method is lacked, and bird-involved fault prevention measures are difficult to take pertinently. Therefore, it is necessary to develop intelligent bird species identification research related to bird-involved faults of overhead transmission lines, and provide basis for line operation and maintenance personnel to correctly identify birds.

At present, the traditional method for identifying the twitter is to extract the characteristics of Linear Prediction Cepstrum Coefficient (LPCC), Mel cepstrum coefficient (MFCC), power spectral density and the like of a sound signal and carry out classification prediction by combining classification algorithms such as Random Forest (RF), Support Vector Machine (SVM), Hidden Markov Model (HMM), Gaussian Mixture Model (GMM) and the like, and the traditional methods are difficult in characteristic extraction and low in identification accuracy.

Disclosure of Invention

Aiming at the problems in the prior art, the invention aims to provide a method for identifying bird-involved fault bird species in a power grid based on Mel language graph and deep learning, which identifies the bird species according to a singing signal and provides a basis for developing targeted and differential bird prevention for a power transmission line.

In order to achieve the purpose, the invention adopts the following technical scheme, which comprises the following steps:

s1: establishing a related bird seed singing database according to the counted main bird seeds with bird-involved faults of the power grid and the actual situation of the power grid;

s2: carrying out denoising, framing and windowing preprocessing on samples in a singing database, removing noise in the singing signal by adopting an improved spectral subtraction method of multi-window spectral estimation, framing the singing signal by setting the frame length and the frame shift, and multiplying the frame length and the frame shift by a window function to increase the continuity of two ends of the frame;

s3: calculating the energy of each frame of the birdsong signal in each Mel filter to obtain the Mel energy of the birdsong sample, obtaining an MxN order matrix containing signal energy information, mapping the energy and the color depth one by one to obtain a Mel language map of the birdsong signal, and dividing the Mel language map into a training set, a verification set and a test set;

s4: building a convolutional neural network classification model, performing repeated iterative training by taking a Mel (Mel) morphogram of a training set as input, testing a verification set in the training process to adjust parameters of the model, and finishing the training when the loss between a predicted output value and an actual value of the network reaches the minimum;

s5: and predicting and identifying the bird species in the test set by using the trained network, and outputting the corresponding bird species category.

Further, the calculation process of the Mel language graph in S3 is as follows: for a section of M-frame birdsong signals, N Mel filters are arranged, an MxN matrix is obtained through Mel energy calculation, a Mel language graph is obtained through coloring according to the energy, horizontal and vertical coordinates in the Mel language graph are the number of frames and the number of the filters respectively, only the MxN data size needs to be calculated, output is simplified, and meanwhile calculation time is shortened.

Further, the convolutional neural network in S4 includes a plurality of convolution-pooling processes for capturing Mel-language graph features, the training set is trained by adjusting network parameters and network iteration times, the model performs prediction on the validation set once and adjusts parameters accordingly according to the prediction result of the validation set every time the model is trained for a certain number of rounds, and the model is corrected in the direction of high prediction accuracy until the loss function value of the network is minimized, and the network training is completed.

The invention has the beneficial effects that:

the method for identifying bird-involved fault bird species of the power grid based on the Mel language graph and the deep learning overcomes the defects of redundant characteristics, large data volume and insufficient discrimination of the traditional voice feature extraction technology, and further promotes accurate bird identification, thereby providing guidance for differential bird prevention and improving the accuracy and effectiveness of bird-involved fault prevention and control of the power transmission line and the transformer substation.

Drawings

FIG. 1 is a flow chart of an implementation of a method for identifying bird species with bird-involved faults in a power grid based on Mel language maps and deep learning;

FIG. 2 is a diagram illustrating a denoising effect of a birdsong signal according to an embodiment of the present invention;

FIG. 3 is a partial birdsong signal waveform and Mel's chart in accordance with an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a convolutional neural network in an embodiment of the present invention.

Detailed Description

The present invention is further described in the following examples, which should not be construed as limiting the scope of the invention, but rather as providing the following examples which are set forth to illustrate and not limit the scope of the invention.

With the rapid development of deep learning, emerging speech recognition methods tend to convert an acoustic signal into a spectrogram such as a chirp spectrogram or a fourier spectrogram as a feature input of a deep learning model. The invention adopts a method of converting a bird song sound signal into a Mel language graph and then combining a convolutional neural network for classification and identification to predict and classify bird species related to bird-related faults of a power transmission line.

The following processes of singing signal processing, Mel language graph calculation and convolutional neural network training of typical bird species related to bird-related faults of the power transmission line are explained in detail, as shown in fig. 1, and the method comprises the following steps:

s1: and establishing a related bird seed singing database according to the statistical main bird species with bird-involved faults of the power grid and the actual condition of the power grid.

In the embodiment, according to the main bird species of the bird-related faults of the power transmission line counted by operation and maintenance personnel of a certain power saving network and by combining the actual situation of the power network, 40 typical birds causing four failure types, namely bird nest type, bird dung type, bird body short-circuit type and bird pecking type, are selected as research objects, and comprise Wu \40491, phoenix-headed wheat chicken, magpie, four-tone rhododendron, night aigren, doodle, swan goose, azalea, bird swallow, dobby, cliff, dycep, spotted dog, ordinary gull, crow, pool aigre, grey-head green pecker, grey-tail 26891bird, grey goose, gray crane, giraffe, white-head spotted dove, white-head bulbul, white aigren, crow, red-mouth gull, red-tail burrower, red-horn, red falcon, belly small owl, wanny duck, cocky bird, cocket, eagle, red-eagle gull, silver carving, chincey, bone-top chicken, quail and a sample database is established and a sample database is disclosed by the birds is collected and public.

S2: the method comprises the steps of carrying out preprocessing such as denoising, framing and windowing on samples in a singing database, removing noise in a singing signal by adopting an improved spectral subtraction method of multi-window spectral estimation, framing the singing signal by setting the frame length and the frame shift, and multiplying the frame length and the frame shift by a window function to increase the continuity of two ends of the frame.

In the embodiment, preprocessing operations such as format unification, denoising, framing, windowing and the like are carried out on all the bird song audio signals, the sampling frequency of all audios is set to 16000Hz by utilizing GoldWave and Sox software, a sound channel is set to be a single sound channel, the length of the audios is uniformly cut to be 1 second, and the audios are stored in a wav format; framing the audio, respectively setting the frame length and the frame shift to be 0.025 second and 0.01 second, and dividing each audio sample into 98 frames; then selecting a Hamming window to carry out windowing operation so as to increase the continuity of two ends of the frame; the speech is denoised by adopting an improved spectral subtraction method of multi-window spectral estimation, and the denoising effect is shown in fig. 2, (a) is the rhododendron speech containing noise, and (b) is the denoised rhododendron speech.

S3: the method comprises the steps of calculating the energy of each frame of the birdsong signal in each Mel filter, calculating the Mel energy of the birdsong sample to obtain an MxN-order matrix containing signal energy information, mapping the energy and the color depth one by one to obtain a Mel spectrogram of the birdsong signal, and dividing the Mel spectrogram into a training set, a verification set and a test set.

The Mel language map is an image representation of the singing signal, and the Mel language maps formed by different bird species have differences. In this embodiment, the bird song signal is divided into 98 frames, 40 Mel filters are provided, a 98 × 40 data matrix is obtained through Mel energy calculation, a Mel language graph of the bird song signal can be obtained through coloring according to the energy, horizontal and vertical coordinates in the Mel language graph are the number of frames and the number of the filters, only 98 × 40 data volumes need to be calculated, output is simplified, and calculation time is shortened. Fig. 3 shows a part of bird song signal waveforms and their corresponding Mel language maps, (a), (b), (c) are the voice waveforms of azalea, red owl and red-bill gull, respectively, (d), (e) and (f) are the Mel language maps of azalea, red-horn owl and red-bill gull, respectively, which describe a section of bird song signal in terms of frame number combined with the number of Mel filters, and can distinguish the song of different bird species.

In this embodiment, the acquired Mel-language graph is divided into a training set, a verification set and a test set according to a ratio of 8: 1.

S4: and (3) building a convolutional neural network classification model, performing repeated iterative training by taking the Mel tone map of the training set as input, testing the verification set in the training process to adjust the parameters of the model, and finishing the training when the loss between the predicted output value and the actual value of the network is minimum.

In this embodiment, a 24-layer convolutional neural network model is constructed, as shown in fig. 4, training is performed with a training set as an input, and the convolutional neural network includes a plurality of convolutions-pooling for capturing Mel-language graph featuresThe process is that an initial learning rate of 0.01 is set to train a training set, the learning rate is reduced to 1/10 after 30 rounds of training, and when a model is trained for a certain round, the model carries out prediction on a verification set once and correspondingly adjusts parameters in a network according to the prediction result of the verification set, and corrects the parameters in the direction with high prediction accuracy. The training of the convolutional neural network is essentially a process of minimizing a loss function, and the aim of learning the optimal class of image feature matching is achieved by continuously iteratively optimizing to seek the minimum loss between the predicted output value and the actual value of the network. The loss function used by the convolutional neural network in this embodiment is a cross entropy function, and the expression is:m is the total number of samples, k is the number of classes of samples, 1{ y }_iJ is an indicative function, the output is 1 when the value in parenthesis is true, otherwise 0,indicating the probability that the ith sample is predicted as the jth class. And when the loss function value of the network is reduced to the minimum, the network training is finished.

S5: and predicting and identifying the bird species in the test set by using the trained network, and outputting the corresponding bird species category.

In this embodiment, since the training samples and the testing samples are randomly selected, in order to avoid the contingency of the classification result, classification tests under 3 groups of different training sample sets are performed, and the average prediction accuracy is 96.7%. Therefore, the Mel spectrogram of the singing signal is used as the characteristic quantity, the voice is denoised by using the improved spectral subtraction of multi-window spectral estimation, and the deep learning model is established by using the convolutional neural network, so that the related bird species threatening the safe operation of the power transmission line can be accurately identified, and guidance is provided for differential bird prevention.

It should be understood that parts of the specification not set forth in detail are well within the prior art.

Although specific embodiments of the present invention have been described above with reference to the accompanying drawings, it will be appreciated by those skilled in the art that these are merely illustrative and that various changes or modifications may be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is only limited by the appended claims.