Transformer operation auxiliary monitoring system based on power grid disaster prevention and reduction system and monitoring method thereof
Technical Field
The invention relates to the field of power grid disaster prevention and reduction, in particular to a transformer operation auxiliary monitoring system and a monitoring method thereof based on a power grid disaster prevention and reduction system.
Background
The power grid disaster prevention and reduction system constructs a power grid three-dimensional model on a three-dimensional geographic information system by constructing a three-dimensional visual simulation platform and defines three-dimensional geographic coordinates. The meteorological bureau data are distributed on each coordinate of power grid equipment and devices in the model, typhoons and raining scenes related to squall lines are visualized, and early warning analysis technologies of windage yaw, loads and the like are realized.
However, the existing disaster prevention and reduction system estimates the natural environment condition of the geographic coordinate point by using the data of the meteorological bureau, and the accuracy is not enough. The position of the electrical device is modeled and the coordinate points have errors. The hysteresis of meteorological data cannot judge the actual conditions of wind speed, rain, temperature and humidity in real time. Moreover, the existing disaster prevention and reduction system has large data of sound information and is limited by poor network transmission bandwidth of remote areas, so that sound signals cannot be transmitted in a short time.
Disclosure of Invention
In view of this, the invention aims to provide a transformer operation auxiliary monitoring system based on a power grid disaster prevention and reduction system and a monitoring method thereof, so that the data volume is reduced, and the identification precision is improved.
The device of the invention is realized by adopting the following scheme: a transformer operation auxiliary monitoring system based on a power grid disaster prevention and reduction system specifically comprises a field data collection module, a data analysis transmission module and a disaster prevention and reduction background module which are connected with one another, wherein the field data collection module comprises a disaster early warning information monitoring unit, and a microclimate monitoring unit, a transformer sound collection unit and a water immersion monitoring unit which are connected with the disaster early warning information monitoring unit; the data analysis and transmission module comprises a sound signal filtering unit, an abnormality judgment unit and a signal transmission unit, wherein the sound signal filtering unit comprises a frequency domain energy detector; the disaster prevention and reduction background module comprises a signal enhancement unit.
The method of the invention is realized by adopting the following scheme: a monitoring method of a transformer operation auxiliary monitoring system based on a power grid disaster prevention and reduction system specifically comprises the following steps:
step S1: after a disaster early warning information monitoring unit in the field data collection module receives an early warning signal sent by a disaster prevention and reduction background module, starting a microclimate monitoring unit to monitor microclimate conditions including ambient temperature, wind speed and humidity, starting a transformer sound collection unit to collect transformer sound, and starting a water immersion monitoring unit to collect water immersion information;
step S2: a sound signal filtering unit in the data analysis transmission module receives the transformer sound collected by the transformer sound collecting unit in the step S1, and filters the transformer sound by a frequency domain energy detection method to judge whether a useful signal exists, if the useful signal exists, the frequency domain energy detector is started to detect the transformer sound or water immersion information, and the process goes to a step S3; if no useful signal exists, returning to the step S1;
step S3: if the frequency domain energy detector detects that the sound or water immersion information of the transformer is abnormal, an abnormality judgment unit is started, and the abnormality judgment unit sequentially performs MFCC feature recognition, establishes a sound recognition probability model of the transformer by adopting a GMM model and judges the sound abnormality of the transformer, further determines and processes the abnormal information, and then the step S4 is carried out; if the frequency domain energy detector detects that the sound and the water immersion information of the transformer are normal, returning to the step S1;
step S4: when the abnormity judgment unit determines that the running sound of the transformer is abnormal, abnormal information is sent to the information transmission unit, the information transmission unit transmits abnormal sound signals collected by the field data collection module to the signal enhancement unit of the disaster prevention and reduction background module, and the signal enhancement unit performs signal enhancement processing on the abnormal sound signals and obtains a transformer abnormity judgment conclusion.
Further, the frequency domain energy detection method in step S2 specifically includes: converting a time domain signal into a frequency domain signal by Fourier transformation of a signal Y (n) to be monitored, squaring the frequency domain signal, accumulating to obtain a frequency domain energy value of the sound signal, and comparing a value obtained by dividing the frequency domain energy value by the noise power with a judgment threshold value lambda so as to judge whether a useful signal exists or not. The method specifically comprises the following steps: the detection threshold λ can be determined by the known false alarm probability PFA when the signal is absent. Because at H0In the case of (2), T follows a Gaussian distribution,thenWherein,therefore, given N and PFA, the detection threshold isWherein the signal-to-noise ratio SNR isFrom the above analysis, it can be seen that, after PFA is determined, the detection threshold λ can be calculated under the condition that the noise variance is known, so that the signal to be monitored can be judged to have a fault sound signal through comparison.
Further, the MFCC feature identification in step S3 specifically includes the following steps;
step S301: passing the frequency filtered sound signal power spectrum through a set of 20 frequency bands of triangular filters, the 20 frequency bands being uniformly distributed over the Mel frequency, wherein the Mel frequency is related to the general frequency f by: mel (f) ═ 2595 × log10(1+ f/700);
step S302, dividing the sound signal into a series of continuous frames, adding Hanning window frames, wherein each frame comprises N512 samples, and 256 samples overlap in adjacent frames; let the audio data time domain signal be x (n), the ith frame audio signal xi(n) can be represented by xi(N) ═ x (i × N + N) w (N),0 ≦ N-1; wherein the Hanning window is:
after Fast Fourier Transform (FFT) is carried out on each frame of signal x (n), a discrete power spectrum P (k) is obtained by taking the square of a module; pi(k) A discrete power spectrum representing the starting position of the ith frequency band, which is exactly the ending position of the (i-2) th frequency band; the output of each triangular filter is obtained:
wherein Wl(k) Is the ith filter, where h (l) represents: h (2/(freq (L +2) -freq (L)), L1, 2,. and L;
step S303: and (3) carrying out logarithmic operation on the output of the filter, and then carrying out discrete cosine change to obtain R Mel frequency cepstrum parameters of each frame, wherein R is 12 to obtain a 12-dimensional feature vector, and the discrete cosine transformation formula is as follows:
wherein t is 1,2, …, n; k is 0,1, …, R; r is more than or equal to 1 and less than or equal to L; n is the number of frames of the sound data file, and L is the number of triangle bands.
Further, the step S3 of establishing the transformer acoustic recognition probability model by using the GMM model specifically includes the following steps:
step S311: assuming that each class of transformer sound is a random feature vector, forming a Gaussian mixture model by M Gaussian components with D dimensions, and using phi for the V-th soundk={μk,ΣkDenotes the parameter of the kth Gaussian component of the model, WkRepresenting the probability of occurrence of the k-th Gaussian component, the weighted sum of the M Gaussian components can be expressed asThe class of sound is represented by a random variable X, and the state value at each time t is the sampled value of the random variable XtMFCC (t, k), the sound probability feature distribution can be expressed as:
wherein mukSum-sigmakRespectively representing an expectation matrix and a covariance matrix, and T represents the total length of the sound signal;
step S312: using GaussAfter the distribution density represents the initial acoustic model, the parameters of the Gaussian mixture model, p, are reestimated by the EM algorithmi(x|φi) Is a Gaussian distribution, pii,μiSum-sigmaiIs the newly estimated parameter value, phihRepresenting old parameter values, p (i | x)l,Φh) Representing the probability that X belongs to the ith distribution, the state value at each time l being the sampled value of a random variable X, the kth state defining XlMFCC (l, k); then
Transformer abnormal sound probability obtained by Bayesian formulaWherein M is a weighted sum of Gaussian components;
step S313: finding out the sound model with the maximum posterior probability by using the known N sampling values and the equations in the steps S311 to S312 and calculating a new parameter value by using the sampling values and the old parameter value.
Further, in step S3, the sound abnormality determination of the transformer specifically includes: according to the conditions of temperature c, humidity h and wind speed s of the field environment, the abnormal judgment probability p is formulatedcThe making process comprises the following steps:
setting environmental factorsWherein c ', h ' and s ' are average temperature, humidity and wind speed in the month;
wherein x is 0.8 is soundIdentifying the probability factor, pmThe failure rate of the transformer is;
when the abnormal sound probability p (i | x) of the transformerl,Φh)<pcAnd prompting that the running sound of the transformer is abnormal and sending abnormal information to the information transmission unit.
Further, the signal enhancement unit in step S4 performs signal enhancement processing on the abnormal sound signal, specifically including the following steps:
step S41: to noise-containing signal power spectrum | F (l, k) converter2Performing first-order recursive smoothing to obtain a smoothed power spectrum Y (l, k): y (l, k) ═ sY (l-1, k) + (1-s) | F (l, k) |2(ii) a Wherein l is a frame number, k is a subband number, sigma is a constant smoothing parameter, and sigma is taken to be 0.7;
step S42: finding the minimum value of Y (l, k) by a forward-backward combined bidirectional search algorithm:
Ymin(l,k)=max{Yf(l,k),Yb(l, k) }; wherein Y isf(l, k) is the minimum value found in the forward direction, Yb(l, k) is the minimum value searched backwards;
step S43: calculating the existence probability p (l, k) of the effective signal: p (l, k) ═ σ1p(l-1,k)+(1-σ1) H (l, k); wherein sigma10.2 is a constant smoothing parameter, and H (l, k) is a criterion for determining the existence of a valid signal, specifically, if Y (l, k)/YminIf (l, k) > phi (k), then H (l, k) ═ 1, which indicates that there is a valid signal in the frame, otherwise H (l, k) > 0, which indicates that there is no valid signal in the frame; phi (k) is a frequency-dependent discrimination threshold, and is 2 when k is less than 1kHz or is between 1 and 3kHz, and is 5 when k is between 3kHz and half of the signal sampling frequency;
step S44: noise power spectrum N (l, k) estimation is performed according to the time-frequency smoothing factor σ (l, k): wherein σ (l, k) ═ σ2+(1-σ2)p(l,k),N(l,k)=σ(l,k)N(l-1,k)+(1-σ(l,k))|F(l,k)|2Where σ is2=0.95,σ2≤σ*(l,k)≤1;
Step S45: calculating a spectral gain factor:
wherein C (l, k) ═ F (l, k) < >2N (l, k) is the clean signal power spectrum, α is the over-subtraction factor, with values:
step S46: the final enhanced signal power spectrum is:
X(l,k)=G(l,k)|F(l,k)|2;
and step S47, the disaster prevention and reduction background module obtains a conclusion of transformer abnormity judgment by comparing the power spectrum of the normal operation sound of the transformer with the current sound power spectrum.
Compared with the prior art, the invention has the following beneficial effects: when a disaster event occurs, whether the transformer normally operates or not is analyzed in real time through the sound information number, whether operation potential safety hazards exist or not, and the device is convenient to install. And on-line monitoring is started according to the early warning information of the power grid disaster prevention and reduction system, the data transmission quantity is small, and the limitation of network bandwidth is small. And the data volume is further reduced and the identification precision is improved by a frequency domain filtering and signal enhancing method.
Drawings
FIG. 1 is a schematic block diagram of the system of the present invention.
Detailed Description
The present invention will be further described with reference to the following examples.
As shown in fig. 1, the present embodiment provides a transformer operation auxiliary monitoring system based on a power grid disaster prevention and reduction system, which specifically includes a field data collection module, a data analysis transmission module, and a disaster prevention and reduction background module that are connected to each other, where the field data collection module includes a disaster early warning information monitoring unit, and a microclimate monitoring unit, a transformer sound collection unit, and a water immersion monitoring unit that are connected to the disaster early warning information monitoring unit; the data analysis and transmission module comprises a sound signal filtering unit, an abnormality judgment unit and a signal transmission unit, wherein the sound signal filtering unit comprises a frequency domain energy detector; the disaster prevention and reduction background module comprises a signal enhancement unit.
The embodiment also provides a monitoring method of the transformer operation auxiliary monitoring system based on the power grid disaster prevention and reduction system, which specifically comprises the following steps:
step S1: after a disaster early warning information monitoring unit in the field data collection module receives an early warning signal sent by a disaster prevention and reduction background module, starting a microclimate monitoring unit to monitor microclimate conditions including ambient temperature, wind speed and humidity, starting a transformer sound collection unit to collect transformer sound, and starting a water immersion monitoring unit to collect water immersion information;
step S2: a sound signal filtering unit in the data analysis transmission module receives the transformer sound collected by the transformer sound collecting unit in the step S1, and filters the transformer sound by a frequency domain energy detection method to judge whether a useful signal exists, if the useful signal exists, the frequency domain energy detector is started to detect the transformer sound or water immersion information, and the process goes to a step S3; if no useful signal exists, returning to the step S1;
step S3: if the frequency domain energy detector detects that the sound or water immersion information of the transformer is abnormal, an abnormality judgment unit is started, and the abnormality judgment unit sequentially performs MFCC feature recognition, establishes a sound recognition probability model of the transformer by adopting a GMM model and judges the sound abnormality of the transformer, further determines and processes the abnormal information, and then the step S4 is carried out; if the frequency domain energy detector detects that the sound and the water immersion information of the transformer are normal, returning to the step S1;
step S4: when the abnormity judgment unit determines that the running sound of the transformer is abnormal, abnormal information is sent to the information transmission unit, the information transmission unit transmits abnormal sound signals collected by the field data collection module to the signal enhancement unit of the disaster prevention and reduction background module, and the signal enhancement unit performs signal enhancement processing on the abnormal sound signals and obtains a transformer abnormity judgment conclusion.
In this embodiment, the method for detecting frequency domain energy in step S2 specifically includes: converting a time domain signal into a frequency domain signal by Fourier transformation of a signal Y (n) to be monitored, squaring the frequency domain signal, accumulating to obtain a frequency domain energy value of the sound signal, and comparing a value obtained by dividing the frequency domain energy value by the noise power with a judgment threshold value lambda so as to judge whether a useful signal exists or not. The method specifically comprises the following steps: the detection threshold λ can be determined by the known false alarm probability PFA when the signal is absent. Because at H0In the case of (2), T follows a Gaussian distribution,thenWherein,therefore, given N and PFA, the detection threshold isWherein the signal-to-noise ratio SNR isFrom the above analysis, it can be seen that when PFA is determined and the noise variance is known, the detection threshold λ can be calculated, and thus the transit ratio can be determinedAnd judging that the signal to be monitored may have a fault sound signal.
In the present embodiment, the MFCC feature recognition in step S3 specifically includes the following steps;
step S301: passing the frequency filtered sound signal power spectrum through a set of 20 frequency bands of triangular filters, the 20 frequency bands being uniformly distributed over the Mel frequency, wherein the Mel frequency is related to the general frequency f by: mel (f) ═ 2595 × log10(1+ f/700);
step S302, dividing the sound signal into a series of continuous frames, adding Hanning window frames, wherein each frame comprises N512 samples, and 256 samples overlap in adjacent frames; let the audio data time domain signal be x (n), the ith frame audio signal xi(n) can be represented by xi(N) ═ x (i × N + N) w (N),0 ≦ N-1; wherein the Hanning window is:
after Fast Fourier Transform (FFT) is carried out on each frame of signal x (n), a discrete power spectrum P (k) is obtained by taking the square of a module; pi(k) A discrete power spectrum representing the starting position of the ith frequency band, which is exactly the ending position of the (i-2) th frequency band; the output of each triangular filter is obtained:
wherein Wl(k) Is the ith filter, where h (l) represents: h (2/(freq (L +2) -freq (L)), L1, 2,. and L;
step S303: and (3) carrying out logarithmic operation on the output of the filter, and then carrying out discrete cosine change to obtain R Mel frequency cepstrum parameters of each frame, wherein R is 12 to obtain a 12-dimensional feature vector, and the discrete cosine transformation formula is as follows:
wherein t is 1,2, …, n; k is 0,1, …, R; r is more than or equal to 1 and less than or equal to L; n is the number of frames of the sound data file, and L is the number of triangle bands.
In this embodiment, the step S3 of establishing the transformer acoustic recognition probability model by using the GMM model specifically includes the following steps:
step S311: assuming that each class of transformer sound is a random feature vector, forming a Gaussian mixture model by M Gaussian components with D dimensions, and using phi for the V-th soundk={μk,ΣkDenotes the parameter of the kth Gaussian component of the model, WkRepresenting the probability of occurrence of the k-th Gaussian component, the weighted sum of the M Gaussian components can be expressed asThe class of sound is represented by a random variable X, and the state value at each time t is the sampled value of the random variable XtMFCC (t, k), the sound probability feature distribution can be expressed as:
wherein mukSum-sigmakRespectively representing an expectation matrix and a covariance matrix, and T represents the total length of the sound signal;
step S312: after representing the initial acoustic model using the Gaussian distribution density, the parameters of the Gaussian mixture model, p, are reestimated by the EM algorithmi(x|φi) Is a Gaussian distribution, pii,μiSum-sigmaiIs the newly estimated parameter value, phihRepresenting old parameter values, p (i | x)l,Φh) Representing the probability that X belongs to the ith distribution, the state value at each time l being the acquisition of a random variable XSample, k-th state defines xlMFCC (l, k); then
Transformer abnormal sound probability obtained by Bayesian formulaWherein M is a weighted sum of Gaussian components;
step S313: finding out the sound model with the maximum posterior probability by using the known N sampling values and the equations in the steps S311 to S312 and calculating a new parameter value by using the sampling values and the old parameter value.
In this embodiment, the sound abnormality determination of the transformer in step S3 specifically includes: according to the conditions of temperature c, humidity h and wind speed s of the field environment, the abnormal judgment probability p is formulatedcThe making process comprises the following steps:
setting environmental factorsWherein c ', h ' and s ' are average temperature, humidity and wind speed in the month;
where x is 0.8, the probability factor for voice recognition, pmThe failure rate of the transformer is;
when the abnormal sound probability p (i | x) of the transformerl,Φh)<pcAnd prompting that the running sound of the transformer is abnormal and sending abnormal information to the information transmission unit.
In this embodiment, the signal enhancement processing of the abnormal sound signal by the signal enhancement unit in step S4 specifically includes the following steps:
step S41: to noise-containing signal power spectrum | F (l, k) converter2Performing first-order recursive smoothing to obtain a smoothed power spectrum Y (l, k): y (l, k) ═ sY (l-1, k) + (1-s) | F (l, k) |2(ii) a Wherein l is a frame number, k is a subband number, sigma is a constant smoothing parameter, and sigma is taken to be 0.7;
step S42: finding the minimum value of Y (l, k) by a forward-backward combined bidirectional search algorithm:
Ymin(l,k)=max{Yf(l,k),Yb(l, k) }; wherein Y isf(l, k) is the minimum value found in the forward direction, Yb(l, k) is the minimum value searched backwards;
step S43: calculating the existence probability p (l, k) of the effective signal: p (l, k) ═ σ1p(l-1,k)+(1-σ1) H (l, k); wherein sigma10.2 is a constant smoothing parameter, and H (l, k) is a criterion for determining the existence of a valid signal, specifically, if Y (l, k)/YminIf (l, k) > phi (k), then H (l, k) ═ 1, which indicates that there is a valid signal in the frame, otherwise H (l, k) > 0, which indicates that there is no valid signal in the frame; phi (k) is a frequency-dependent discrimination threshold, and is 2 when k is less than 1kHz or is between 1 and 3kHz, and is 5 when k is between 3kHz and half of the signal sampling frequency;
step S44: noise power spectrum N (l, k) estimation is performed according to the time-frequency smoothing factor σ (l, k): wherein σ (l, k) ═ σ2+(1-σ2)p(l,k),N(l,k)=σ(l,k)N(l-1,k)+(1-σ(l,k))|F(l,k)|2Where σ is2=0.95,σ2≤σ*(l,k)≤1;
Step S45: calculating a spectral gain factor:
wherein C (l, k) ═ F (l, k) < >2N (l, k) is the clean signal powerSpectrum, α is an over-subtraction factor with values:
step S46: the final enhanced signal power spectrum is:
X(l,k)=G(l,k)|F(l,k)|2;
and step S47, the disaster prevention and reduction background module obtains a conclusion of transformer abnormity judgment by comparing the power spectrum of the normal operation sound of the transformer with the current sound power spectrum.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.