CN113724726A - Unit operation noise suppression processing method based on full-connection neural network - Google Patents
Unit operation noise suppression processing method based on full-connection neural network Download PDFInfo
- Publication number
- CN113724726A CN113724726A CN202110950443.7A CN202110950443A CN113724726A CN 113724726 A CN113724726 A CN 113724726A CN 202110950443 A CN202110950443 A CN 202110950443A CN 113724726 A CN113724726 A CN 113724726A
- Authority
- CN
- China
- Prior art keywords
- neural network
- fully
- connected neural
- acoustic
- normalization
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 53
- 238000010977 unit operation Methods 0.000 title claims abstract description 21
- 238000003672 processing method Methods 0.000 title claims abstract description 15
- 230000001629 suppression Effects 0.000 title description 5
- 239000002184 metal Substances 0.000 claims abstract description 24
- 238000000034 method Methods 0.000 claims abstract description 14
- 230000002401 inhibitory effect Effects 0.000 claims abstract description 8
- 238000010606 normalization Methods 0.000 claims description 31
- 238000005070 sampling Methods 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 18
- 238000012545 processing Methods 0.000 claims description 11
- 238000012795 verification Methods 0.000 claims description 8
- 230000009466 transformation Effects 0.000 claims description 4
- 230000002159 abnormal effect Effects 0.000 abstract description 6
- 238000012544 monitoring process Methods 0.000 abstract description 4
- 230000005236 sound signal Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 6
- 238000010079 rubber tapping Methods 0.000 description 3
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L21/0216—Noise filtering characterised by the method used for estimating noise
- G10L21/0232—Processing in the frequency domain
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- Signal Processing (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a processing method for suppressing unit operation noise based on a fully-connected neural network. Compared with the traditional method, the fully-connected neural network is used for inhibiting the unit operation noise, the signal-to-noise ratio of the abnormal sound signal is obviously improved, the faults such as metal collision and the like generated when the hydraulic generator set operates can be identified, and the method can be applied to acoustic monitoring of the operation of the hydraulic generator set; the corresponding acoustic samples can be automatically generated for suppressing and reducing noise through noise identification.
Description
Technical Field
The invention belongs to the field of on-line monitoring of operation of hydroelectric generating set equipment, and particularly relates to a processing method for suppressing unit operation noise based on a fully-connected neural network.
Background
Under the acoustic monitoring environment when the hydroelectric generating set operates, the requirement on the signal to noise ratio of an abnormal acoustic signal is high; because the background noise is large and the frequency domain range is wide when the unit operates, the signal-to-noise ratio condition processed by the traditional signal processing method is not ideal; after signal processing is carried out by utilizing the fully-connected neural network, unit operation noise can be well inhibited, and clear abnormal sound signals such as knocking, collision and the like are obtained; therefore, the abnormal state of the equipment generated when the unit operates can be alarmed; therefore, a processing method for suppressing the unit operation noise based on the fully-connected neural network is needed to be designed.
Disclosure of Invention
Compared with the traditional method, the full-connection neural network is used for suppressing the unit operation noise, the signal-to-noise ratio of the abnormal sound signal is obviously improved, faults such as metal collision and the like generated when the hydroelectric generating set operates can be identified, and the corresponding sound sample is generated according to identification to suppress and reduce the noise.
In order to realize the technical effects, the technical scheme adopted by the invention is as follows: a processing method for suppressing unit operation noise based on a fully-connected neural network comprises the following steps:
s1, training the fully-connected neural network: by collecting sound samples, a fully connected neural network is created that contains two hidden layers:
and S2, collecting and preparing a sound sample:
a. the microphone collects an acoustic signal with a sampling rate of 64KHz, and the signal is down-sampled to a sampling rate of 16 KHz;
b. carrying out short-time Fourier transform on the signals by taking the window length as 256 points and the overlapping rate as 75 percent and taking the window shape as a Hamming window to form an array; carrying out normalization processing on the array according to the requirement of the neural network;
s3, applying a neural network to reduce noise:
a. inputting the array after the normalization processing to a trained fully-connected neural network for prediction to obtain a new array, and performing inverse normalization on the array;
b. and carrying out short-time Fourier inverse transformation on the new array according to the window length of 256 points and the overlapping rate of 75% to obtain an acoustic signal which is used for inhibiting the unit operation noise and has the sampling rate of 16 KHz.
Preferably, the training method of the fully-connected neural network in step S1 is as follows:
s101, collecting an acoustic sample of metal collision fault of equipment when a unit is stopped in a unit stop state; the acoustic sample is down-sampled to a sampling rate of 16KHz, the data is processed by short-time Fourier transform to obtain a two-dimensional array with the window length of 256 points and the overlapping rate of 75 percent and the window shape of a Hamming window, and the two-dimensional array is subjected to normalization operation and normalized parameters are recorded and used as a target set of a training network;
s201, collecting a background noise sample when the water-turbine generator set runs in a normal running state of the water-turbine generator set; randomly superposing a metal collision fault acoustic sample collected before with a running background noise sample of the water-turbine generator set; the method comprises the steps of sampling an acoustic sample after superposition to a sampling rate of 16KHz, carrying out short-time Fourier transform on data with a window length of 256 points and an overlap rate of 75% and a window shape of a Hamming window to obtain a two-dimensional array, carrying out normalization operation on the two-dimensional array, and recording normalization parameters to serve as a prediction set of a training network;
s301, enabling the acoustic samples corresponding to the target set and the prediction set to correspond one to one, and extracting 10% of the acoustic samples to serve as a verification set;
s401, inputting a target set, a prediction set and a verification set into a fully-connected neural network comprising two hidden layers, and repeating the training for 3-5 times;
and S501, recording the structure and parameters of the fully-connected neural network to obtain the fully-connected neural network for inhibiting group operation noise.
Preferably, the normalization operations in S101 and S201 use L2 regularization as a normalization method.
Preferably, the acoustic sample of the metal collision fault comprises acoustic samples of knocking and collision of various metal parts at different parts of the unit.
The invention has the beneficial effects that:
the invention discloses a training method of a unit operation noise suppression fully-connected neural network, which trains a fully-connected neural network comprising two hidden layers by using an acoustic sample of a typical metal collision fault of equipment when a water-turbine generator unit is shut down and a background noise sample of the unit in operation so as to suppress the background noise of the unit in normal operation. Compared with the traditional method, the fully-connected neural network is used for inhibiting the unit operation noise, the signal-to-noise ratio of the abnormal sound signal is obviously improved, the faults such as metal collision and the like generated when the hydraulic generator set operates can be identified, and the method can be applied to acoustic monitoring of the operation of the hydraulic generator set; the corresponding acoustic samples can be automatically generated for suppressing and reducing noise through noise identification.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a block diagram of a process for fully-connected neural network training in accordance with the present invention;
FIG. 3 is a time domain diagram of a typical part tapping sound sample during a unit shutdown condition in accordance with an embodiment of the present invention;
FIG. 4 is a time domain diagram of a section of background noise sample in a wind tunnel during normal operation of a unit in an embodiment of the present invention;
FIG. 5 is a time domain diagram of a typical metal tapping sound sample after a section of background noise of the unit is superimposed in an embodiment of the invention;
FIG. 6 is a time domain diagram of typical part knocking and colliding sounds after noise reduction of a fully connected neural network.
Detailed Description
Example 1:
as shown in fig. 1, a processing method for suppressing unit operation noise based on a fully-connected neural network includes the following steps:
s1, training the fully-connected neural network: by collecting sound samples, a fully connected neural network is created that contains two hidden layers:
and S2, collecting and preparing a sound sample:
a. the microphone collects an acoustic signal with a sampling rate of 64KHz, and the signal is down-sampled to a sampling rate of 16 KHz;
b. carrying out short-time Fourier transform on the signals by taking the window length as 256 points and the overlapping rate as 75 percent and taking the window shape as a Hamming window to form an array; carrying out normalization processing on the array according to the requirement of the neural network;
s3, applying a neural network to reduce noise:
a. inputting the array after the normalization processing to a trained fully-connected neural network for prediction to obtain a new array, and performing inverse normalization on the array;
b. and carrying out short-time Fourier inverse transformation on the new array according to the window length of 256 points and the overlapping rate of 75% to obtain an acoustic signal which is used for inhibiting the unit operation noise and has the sampling rate of 16 KHz.
Preferably, the training method of the fully-connected neural network in step S1 is as follows:
s101, collecting an acoustic sample of metal collision fault of equipment when a unit is stopped in a unit stop state; the acoustic sample is down-sampled to a sampling rate of 16KHz, the data is processed by short-time Fourier transform to obtain a two-dimensional array with the window length of 256 points and the overlapping rate of 75 percent and the window shape of a Hamming window, and the two-dimensional array is subjected to normalization operation and normalized parameters are recorded and used as a target set of a training network;
s201, collecting a background noise sample when the water-turbine generator set runs in a normal running state of the water-turbine generator set; randomly superposing a metal collision fault acoustic sample collected before with a running background noise sample of the water-turbine generator set; the method comprises the steps of sampling an acoustic sample after superposition to a sampling rate of 16KHz, carrying out short-time Fourier transform on data with a window length of 256 points and an overlap rate of 75% and a window shape of a Hamming window to obtain a two-dimensional array, carrying out normalization operation on the two-dimensional array, and recording normalization parameters to serve as a prediction set of a training network;
s301, enabling the acoustic samples corresponding to the target set and the prediction set to correspond one to one, and extracting 10% of the acoustic samples to serve as a verification set;
s401, inputting a target set, a prediction set and a verification set into a fully-connected neural network comprising two hidden layers, and repeating the training for 3-5 times;
and S501, recording the structure and parameters of the fully-connected neural network to obtain the fully-connected neural network for inhibiting group operation noise.
Preferably, the normalization operations in S101 and S201 use L2 regularization as a normalization method.
Preferably, the acoustic sample of the metal collision fault comprises acoustic samples of knocking and collision of various metal parts at different parts of the unit.
Example 2:
the fully-connected neural network is trained as shown in fig. 2:
s1, collecting acoustic samples of a plurality of typical metal knocking and colliding wind shields, stator cores and rotors under the shutdown state of the unit by using a hydraulic generator noise monitor with the sampling rate of 64KHz and the sampling precision of 16bit, wherein the acoustic samples are 23 sections in total, and the sampling rate of the acoustic samples is reduced to 16 KHz; the tapping sample is shown in fig. 2.
And S2, acquiring background noise in the wind tunnel when the unit normally operates by using the hydraulic generator noise monitor at a sampling rate of 16bit of 64KHz, extracting 4 sections of background noise samples of the unit under active power working conditions of 147MW, 130MW, 110MW and 100MW, and reducing the sampling rate of the sound samples to 16 KHz. Here, a background noise sample of a section of the unit operation is shown, and a knock sample is shown in fig. 3.
And S3, combining and superposing the unit background noise sample and the typical metal knocking and collision sound sample to obtain 92 sections of typical metal knocking and collision sound samples with background noise. Here a certain section of the superimposed sound sample is shown, the tap sample being shown in fig. 3.
S4, carrying out short-time Fourier transform on the superposed sound samples with the window length of 256 points and the overlapping rate of 75%, and carrying out normalization processing on the transform results to obtain a two-dimensional array as a prediction set of the fully-connected neural network; using L2 regularization as a normalization method with the parameters input mean 15.0355 and input standard deviation 73.0891; and performing the same Fourier transform and normalization processing on the original knocking sound samples and the original collision sound samples without the superimposed background noise to obtain a two-dimensional array serving as a target set of the fully-connected neural network.
S5, a fully-connected neural network with two hidden layers is created.
S6, the prediction set is in one-to-one correspondence with the target set samples, and part of the prediction set is extracted to be used as a test verification set; inputting the prediction set, the target set and the verification set into a fully-connected neural network for training, repeating 3 rounds, and finally obtaining a mean square error RMSE value of 6.7257.
And S7, storing the trained network structure, weight parameters and normalization parameters.
Real-time example 3:
as shown in fig. 3 to 6:
s1, carrying out short-time Fourier transform on typical metal knocking and collision sounds superposed with unit background noises with the window length of 256 points and the overlapping rate of 75%, carrying out normalization processing on the transform result, and inputting the obtained two-dimensional array into a trained fully-connected neural network for prediction;
s2, carrying out inverse normalization processing on the array output by the fully-connected neural network according to the previously stored normalization parameters to obtain a new array;
s3, carrying out short-time Fourier inverse transformation on the new array according to the window length of 256 points and the overlapping rate of 75% to obtain an acoustic signal with the sampling rate of 16KHz, wherein the sound is typical metal knocking and collision sound after the unit operation noise is suppressed; the background noise suppression effect is shown in fig. 6; it can be seen that the typical metal knocking and collision sound waveforms obtained by the method after the unit background noise is suppressed are consistent with the original metal knocking and collision sound waveforms without the superimposed background noise, and the unit background noise suppression effect is good.
The above-described embodiments are merely preferred embodiments of the present invention, and should not be construed as limiting the present invention, and features in the embodiments and examples in the present application may be arbitrarily combined with each other without conflict. The protection scope of the present invention is defined by the claims, and includes equivalents of technical features of the claims. I.e., equivalent alterations and modifications within the scope hereof, are also intended to be within the scope of the invention.
Claims (4)
1. A processing method for suppressing unit operation noise based on a fully-connected neural network is characterized by comprising the following steps: it comprises the following steps:
s1, training the fully-connected neural network: by collecting sound samples, a fully connected neural network is created that contains two hidden layers:
and S2, collecting and preparing a sound sample:
a. the microphone collects an acoustic signal with a sampling rate of 64KHz, and the signal is down-sampled to a sampling rate of 16 KHz;
b. carrying out short-time Fourier transform on the signals by taking the window length as 256 points and the overlapping rate as 75 percent and taking the window shape as a Hamming window to form an array; carrying out normalization processing on the array according to the requirement of the neural network;
s3, applying a neural network to reduce noise:
a. inputting the array after the normalization processing to a trained fully-connected neural network for prediction to obtain a new array, and performing inverse normalization on the array;
b. and carrying out short-time Fourier inverse transformation on the new array according to the window length of 256 points and the overlapping rate of 75% to obtain an acoustic signal which is used for inhibiting the unit operation noise and has the sampling rate of 16 KHz.
2. The processing method for suppressing the unit operation noise based on the fully-connected neural network according to claim 1, wherein the processing method comprises the following steps: the training method of the fully-connected neural network in step S1 is as follows:
s101, collecting an acoustic sample of metal collision fault of equipment when a unit is stopped in a unit stop state; the acoustic sample is down-sampled to a sampling rate of 16KHz, the data is processed by short-time Fourier transform to obtain a two-dimensional array with the window length of 256 points and the overlapping rate of 75 percent and the window shape of a Hamming window, and the two-dimensional array is subjected to normalization operation and normalized parameters are recorded and used as a target set of a training network;
s201, collecting a background noise sample when the water-turbine generator set runs in a normal running state of the water-turbine generator set; randomly superposing a metal collision fault acoustic sample collected before with a running background noise sample of the water-turbine generator set; the method comprises the steps of sampling an acoustic sample after superposition to a sampling rate of 16KHz, carrying out short-time Fourier transform on data with a window length of 256 points and an overlap rate of 75% and a window shape of a Hamming window to obtain a two-dimensional array, carrying out normalization operation on the two-dimensional array, and recording normalization parameters to serve as a prediction set of a training network;
s301, enabling the acoustic samples corresponding to the target set and the prediction set to correspond one to one, and extracting 10% of the acoustic samples to serve as a verification set;
s401, inputting a target set, a prediction set and a verification set into a fully-connected neural network comprising two hidden layers, and repeating the training for 3-5 times;
and S501, recording the structure and parameters of the fully-connected neural network to obtain the fully-connected neural network for inhibiting group operation noise.
3. The processing method for suppressing the unit operation noise based on the fully-connected neural network according to claim 2, wherein the processing method comprises the following steps: the normalization operations in S101 and S201 use L2 regularization as the normalization method.
4. The processing method for suppressing the unit operation noise based on the fully-connected neural network according to claim 2, wherein the processing method comprises the following steps: the acoustic samples of the metal collision faults comprise acoustic samples of various metal parts knocked and collided at different parts of the unit.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110950443.7A CN113724726A (en) | 2021-08-18 | 2021-08-18 | Unit operation noise suppression processing method based on full-connection neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110950443.7A CN113724726A (en) | 2021-08-18 | 2021-08-18 | Unit operation noise suppression processing method based on full-connection neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113724726A true CN113724726A (en) | 2021-11-30 |
Family
ID=78676741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110950443.7A Pending CN113724726A (en) | 2021-08-18 | 2021-08-18 | Unit operation noise suppression processing method based on full-connection neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113724726A (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859767A (en) * | 2019-03-06 | 2019-06-07 | 哈尔滨工业大学（深圳） | A kind of environment self-adaption neural network noise-reduction method, system and storage medium for digital deaf-aid |
| CN112857767A (en) * | 2021-01-18 | 2021-05-28 | 中国长江三峡集团有限公司 | Hydro-turbo generator set rotor fault acoustic discrimination method based on convolutional neural network |
-
2021
- 2021-08-18 CN CN202110950443.7A patent/CN113724726A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859767A (en) * | 2019-03-06 | 2019-06-07 | 哈尔滨工业大学（深圳） | A kind of environment self-adaption neural network noise-reduction method, system and storage medium for digital deaf-aid |
| CN112857767A (en) * | 2021-01-18 | 2021-05-28 | 中国长江三峡集团有限公司 | Hydro-turbo generator set rotor fault acoustic discrimination method based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740523B (en) | Power transformer fault diagnosis method based on acoustic features and neural network | |
| WO2022261805A1 (en) | Diesel engine gearbox fault diagnosis method | |
| CN108168891B (en) | Method and equipment for extracting weak fault signal characteristics of rolling bearing | |
| CN110534118A (en) | Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network | |
| CN104677623B (en) | A kind of blade of wind-driven generator fault acoustics in place diagnostic method and monitoring system | |
| CN112857767B (en) | Hydro-turbo generator set rotor fault acoustic discrimination method based on convolutional neural network | |
| CN106546892A (en) | The recognition methodss of shelf depreciation ultrasonic audio and system based on deep learning | |
| CN108490349A (en) | Motor abnormal sound detection method based on Mel frequency cepstral coefficients | |
| CN106771598B (en) | A kind of Adaptive spectra kurtosis signal processing method | |
| CN105698922A (en) | Voltage transformer vibration fault feature extraction method based on improved EMD method and Spectral Kurtosis method | |
| CN112329914B (en) | Fault diagnosis method and device for buried transformer substation and electronic equipment | |
| CN112599134A (en) | Transformer sound event detection method based on voiceprint recognition | |
| Li et al. | Weak crack detection for gearbox using sparse denoising and decomposition method | |
| CN107886078A (en) | A kind of Threshold Denoising method based on layered self-adapting threshold function table | |
| CN110909827A (en) | Noise reduction method suitable for fan blade sound signals | |
| CN108036940A (en) | A kind of Method for Bearing Fault Diagnosis | |
| CN108869145A (en) | Pumping plant unit diagnostic method based on compound characteristics index and depth limit learning machine | |
| CN113790911B (en) | Abnormal sound detection method based on sound spectrum statistics rule | |
| Geng et al. | Mechanical fault diagnosis of power transformer by GFCC time-frequency map of acoustic signal and convolutional neural network | |
| Yao et al. | Multiband weights-induced periodic sparse representation for bearing incipient fault diagnosis | |
| CN113724726A (en) | Unit operation noise suppression processing method based on full-connection neural network | |
| Yao et al. | Radiation Noise Separation of Internal Combustion Engine Based on Gammatone‐RobustICA Method | |
| CN110412366B (en) | Dynamic time algorithm-based transformer live-line detection method | |
| Yang et al. | Iterative Morlet wavelet with SOSO boosting strategy for impulsive feature extraction | |
| Li et al. | Envelope analysis by wavelet-filter based spectral kurtosis for bearing health monitoring |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |