CN107609488A - A kind of ship noise method for identifying and classifying based on depth convolutional network - Google Patents
A kind of ship noise method for identifying and classifying based on depth convolutional network Download PDFInfo
- Publication number
- CN107609488A CN107609488A CN201710717199.3A CN201710717199A CN107609488A CN 107609488 A CN107609488 A CN 107609488A CN 201710717199 A CN201710717199 A CN 201710717199A CN 107609488 A CN107609488 A CN 107609488A
- Authority
- CN
- China
- Prior art keywords
- node
- noise
- layer
- value
- mfcc
- 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.)
- Granted
Links
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a kind of ship noise method for identifying and classifying based on depth convolutional network, belongs to the field that underwater ship noise identifies.The present invention is few primarily directed to the quantity of documents of BP neural network processing, the solution extracted feature unobvious and proposed the problem of being easily absorbed in locally optimal solution.The invention first removes the noise in original sound according to MFCC, extracts corresponding effective feature, and this process is mainly to remove the big noise of interference.Audio files by MFCC processing is converted into the form that depth convolutional network can receive.By the effective feature of the multi-level extraction of depth convolutional neural networks, the validity for the feature so extracted is stronger, with more universality.Classification is identified for a large amount of voice datas in reality, reduces artificial intervention degree, can preferably distinguish different ship noises, so as to reach the purpose of identification.
Description
Technical field
The present invention relates to a kind of ship noise method for identifying and classifying based on depth convolutional network, belong to underwater ship noise
The field of identification.
Background technology
Ship noise, is divided into airborne noise (the caused noise in air dielectric) and water noise (produces in aqueous medium
Noise) wherein naval vessel water noises are divided into ship-radiated noise and naval vessel self noise ship-radiated noises are by mechanical on naval vessel
Operating and Ship Motion produce and are radiated the noise in water, and ship-radiated noise is the information source of passive detection device, is warship
One of concealed important indicator of ship.And mechanical noise and propeller noise constitute main radiated noise, both noises
Importance, depending on the speed of a ship or plane and depth on naval vessel and the frequency of noise, and hydrodynamic noise self noise is influenceed it is big.It is logical
Often, cavitation once occurs for propeller, just often turns into Main Noise Sources, particularly high band.In the low speed of a ship or plane, mechanical noise is past
It is past to rise to Main Noise Sources.In a general case, radiated noise is made up of broadband continuous spectrum and a series of line spectrums.Its center line
Spectrum part and propulsion system, propeller and subsidiary engine are relevant.The typically stabilization of line spectrum component caused by subsidiary engine, and this kind of line spectrum
Amplitude and frequency have periodically variable modulation phenomenon as the speed of a ship or plane on naval vessel changes and changes.Although all kinds of naval vessels itself
And residing condition is different, radiated noise characteristic can change, but similar ship noise characteristic always has certain similitude, without
The ship noise characteristic of same type then has certain difference, this utilization Noise Identification submarine target made it possible.
Due to the complexity of marine environment, it is most important that Classification and Identification is carried out to naval vessel using naval vessel noise.Before
Have in many articles and Classification and Identification is carried out using BP neural network.But because BP neural network only has three layers, so can not
It is complete to extract effective feature.Simultaneously BP neural network for data volume it is fewer when use.It is substantial amounts of when having
When data produce, the recognition effect of BP neural network can be decreased obviously.Most importantly, it is easy to make it using BP neural network
It is absorbed in locally optimal solution.
In the production of reality, substantial amounts of audio files may be handled, and the difference between sound is not
It is too big.Therefore just become extremely difficult using Processing with Neural Network file, and the difference characteristic of three-layer network extraction also can
It is less obvious, it so can then cause identification chaotic so that recognition effect becomes excessively poor.
The content of the invention
It is an object of the invention to provide a kind of ship noise method for identifying and classifying based on depth convolutional network.
The purpose of the present invention through the following steps that realize：
A kind of ship noise method for identifying and classifying based on depth convolutional network, it is characterised in that comprise the steps of：
Step 1 extracts original ship noise；And noise is split, noise is cut into 100ms mono-；It is sent into
Handled in MFCC model；
Step 2 by MFCC handle caused by ship noise matrix be stored in the file specified；
After step 3 is handled by MFCC, acquisition value is the matrix between [- 1,1]；Each value is multiplied by
100000, for negative value, after taking absolute value, 100000 are being multiplied by, is changing into the relatively obvious gray-scale map of feature；
Gray-scale map is carried out the processing that labels by step 4, then, by the use of gray-scale map sum 70% as training set, uses it
Remaining 30% is used as test set；
The gray-scale map of classification is converted to LDMB forms by step 5；
Step 6 by LDMB forms to gray-scale map be sent in depth network, tested by LeNet network trainings, point
Analyse result.
Described MFCC processing, is comprised the steps of：
Step 1 carries out preemphasis to the data of reading first, by a high-pass filter, the purpose of preemphasis be to
Different frequency range adds weights, improves HFS, makes the frequency spectrum of signal become flat；Formula is as follows：
H (Z)=1- μ Z-1 (1)
μ value is between 0.9-1.0 in formula 1；
N number of sampling point set is first synthesized an observation unit, referred to as frame by step 2；Under normal circumstances N value be 256 or
512, the time covered is about 20~30ms or so；In order to avoid the change of adjacent two frame is excessive, allowing between two consecutive frames has one
Section overlapping region, this overlapping region contain M sample point, and M value is about the 1/2 or 1/3 of N；
Step 3 uses Hamming window, to increase the continuity on the frame left side and the right in this stage of adding window；
Signal is transformed from the time domain to frequency domain by step 4 by fast Fourier；
Triangle filtering group of the step 5 by energy spectrum by one group of MEL yardstick, the number of triangular filter are typically chosen
18-22 optimal；
Step 6 obtains MFCC coefficient by discrete cosine transform.
Described LeNet networks, mainly comprising convolutional layer, lower sampling layer, connect layer entirely；The 5*5 sizes that convolutional layer uses
Convolution kernel, and convolution kernel slides a pixel every time, a characteristic spectrum uses same convolution kernel, in the value of each upper layer node
Parameter value, these products and an offset parameter are added to obtain an output, this be input to one activation letter
Number, the output of activation primitive is the value of next node；Lower sampling layer is using 2*2 input domain, 4 sections of last layer
Input of the point as next layer of a node, and input domain is not overlapping, i.e., slides 2 pixels every time；The 4 of each lower sampling node
It is averaged after the summation of individual input node, average is multiplied by a parameter plus an offset parameter as the defeated of activation primitive
Enter, the output of activation primitive is the input of next node layer；Sampling is only with two training parameters under one；Full articulamentum, i.e.,
Some node of a certain layer and each node of last layer connect, and each node each uses set of parameter, what is connected entirely
In network, if K layers have n node, K+1 layers have m node；A then shared n*m connection, has plus each K+1 node layers outside
One biasing, then share n*m+m training parameter.
Beneficial effects of the present invention are：
The present invention, by MFCC processing, some interference noises is got rid of when extracting noise characteristic.The spy so retained
Sign is nearly all validity feature.By the matrix conversion of these validity features into gray-scale map.Carried out in depth convolutional neural networks
Learning classification.98% discrimination is reached.However, for traditional neutral net, the extraction effect of feature be not it is too obvious,
So as to cause identification chaotic, the effect of identification is not reached.This time the model of invention has universality.It can be directed to big in reality
Classification is identified in amount voice data.And artificial intervention degree also relative reduction.So that the effect more authenticity of training.
Brief description of the drawings
Fig. 1 is the operational flowchart of the present invention；
Fig. 2 is the feature extraction flow of MFCC models；
Fig. 3 is the model schematic of convolutional layer；
Fig. 4 is the model schematic of sampling layer；
Fig. 5 is the model schematic of full articulamentum.
Embodiment
The method carried below in conjunction with the accompanying drawings to the present invention is further elaborated：
At present, all it is extracted perhaps in positive research, scholar for submarine target Noise Identification, the expert in domestic and international field
More models and method.Want to find a blanket model, and do not needed too much during processing
Human intervention.However, for existing model, it is extremely difficult at this 2 points, for BP, this model discomfort use
The audio files of big data quantity is managed, and network layer, than shallower, extraction feature is not complete enough.Come for some other models
Say the intervention, it is necessary to artificial.The influence of subjective factor is bigger.And this model is by analysis conventional model, according to conventional model
Weak point be improved, it is proposed that the model algorithm based on MFCC and depth convolution.
The quantity of documents that this invention is handled primarily directed to BP neural network is few, extracts feature unobvious and is easily absorbed in
The solution that the problem of locally optimal solution proposes.The invention first removes the noise in original sound according to MFCC, extraction
Go out corresponding effective feature.This process is mainly to remove the big noise of interference.Audio files by MFCC processing is turned
Change the form that depth convolutional network can receive into.By the effective feature of the multi-level extraction of depth convolutional neural networks, this
The validity of the feature of sample extraction is stronger.Different ship noises can be preferably distinguished, so as to reach the purpose of identification.
Its main points of view and content are as follows：
(1) voice data is pre-processed by MFCC, MFCC is generally used for the extraction to speech sound feature.
By MFCC with identifying it is once new trial in ship noise.During noise characteristic being extracted in MFCC,
We add the Duplication of each frame, can so increase the continuity between each frame.
(2) after being handled by MFCC, the value of the matrix of acquisition is between [- 1,1], and the gray-scale map so changed into is
Black, feature difference unobvious.Each value is multiplied by 100000, for negative value, after taking absolute value, is being multiplied by 100000.
Matrix can thus be changed into the relatively obvious gray-scale map of feature.
(3) for gray-scale map, the processing that labels is carried out, then, the gray-scale map after processing is sent to depth convolution networking
In be trained classification.In this experiment, the network of use is LeNet networks.
It is as shown in Figure 2 in the process that MFCC processing early stage is carried out to naval vessel：
(1) first, preemphasis is carried out to the data of reading, the purpose of preemphasis is to add weights to different frequency range, is improved
HFS, makes the frequency spectrum of signal become flat.Preemphasis is as follows by a high-pass filter, formula：
H (Z)=1- μ Z-1 (1)
μ value is between 0.9-1.0, in this invention, using 0.97 in formula 1.
(2) N number of sampling point set is first synthesized into an observation unit, referred to as frame.N value is 256 or 512 under normal circumstances,
The time covered is about 20~30ms or so.In order to avoid the change of adjacent two frame is excessive, therefore it can allow between two consecutive frames and have
One section of overlapping region, this overlapping region contain M sample point, and usual M value is about the 1/2 or 1/3 of N.What is specifically invented
Used N values are 512, M 384, sample frequency 65531HZ.
(3) in this stage of adding window, Hamming window is used, to increase the continuity on the frame left side and the right.
(4) Fast Fourier Transform (FFT) is a MFCC important step, because change of the signal in time domain is difficult to see
Go out the characteristic of signal.Therefore, the distribution for the energy being generally converted on frequency domain is observed, and the characteristic of different energy represents
Different characteristics of speech sounds.Therefore, it is necessary to time-domain signal is converted into by frequency-region signal by Fast Fourier Transform (FFT).
(5) the triangle filtering group by energy spectrum by one group of MEL yardstick, the number of triangular filter typically choose 18-22
It is individual optimal, in this invention, using 18.By filtering, frequency spectrum can be smoothed, the work of harmonic carcellation
With highlighting the formant of original noise.
(6) MFCC coefficient is obtained by discrete cosine transform.
Because the noise characteristic obtained by MFCC can only represent static nature.But handle is association of activity and inertia and can more could embodied
The feature of ship noise, therefore, first-order difference is introduced to represent the behavioral characteristics of ship noise.
LeNet is shared 7 layers (not including input layer), and every layer all includes different training parameters.Mainly have convolutional layer, under take out
Sample layer, connect connected mode in layer 3 entirely.
Convolutional layer is all the convolution kernel of the 5*5 sizes used, and convolution kernel slides a pixel every time, and a characteristic spectrum makes
With same convolution kernel, the structure of convolution kernel as shown in figure 3, parameter value in the value of each upper layer node, these products and
One offset parameter is added to obtain an output, this is input to an activation primitive, the output of activation primitive is
The value of next node.Convolution kernel has 25 Connecting quantities to bias totally 26 training parameters plus one.
Lower sampling layer is using 2*2 input domain, and 4 nodes of last layer are as the defeated of next layer of a node
Enter, and input domain is not overlapping, i.e., slides 2 pixels every time.The structure of lower sampling node is as shown in figure 4, each descend the 4 of sampling node
It is averaged after the summation of individual input node, average is multiplied by a parameter plus an offset parameter as the defeated of activation primitive
Enter, the output of activation primitive is the input of next node layer.Sampling is only with two training parameters under one.
Full articulamentum, i.e., some node of a certain layer and each node of last layer connect, and each node each uses
Set of parameter, in the network connected entirely, it is assumed that K layers have n node, and K+1 layers have m node.A then shared n*m connection,
There is a biasing plus each K+1 node layers outside, then share n*m+m training parameter, the network structure connected entirely such as Fig. 5 institutes
Show.
A kind of model of the underwater ship noise identification based on MFCC+ depth convolutional networks, concrete implementation mode are as follows
It is shown：
Step 1：Extract original ship noise.And noise is split, noise is cut into 100ms mono-.It is sent into
MFCC model in handle.
Step 2：By MFCC handle caused by ship noise matrix be stored in the file specified.
Step 3：The ship noise eigenmatrix of preservation is converted into gray-scale map.
Step 4：Gray-scale map is labelled and trained, we use it by the use of gray-scale map sum 70% as training set
Remaining 30% is used as test set.
Step 5：The result of test is analyzed.
Claims (3)
1. a kind of ship noise method for identifying and classifying based on depth convolutional network, it is characterised in that comprise the steps of：
Step 1 extracts original ship noise；And noise is split, noise is cut into 100ms mono-；It is sent into MFCC
Model in handle；
Step 2 by MFCC handle caused by ship noise matrix be stored in the file specified；
After step 3 is handled by MFCC, acquisition value is the matrix between [- 1,1]；Each value is multiplied by 100000, it is right
In negative value, after taking absolute value, 100000 are being multiplied by, is changing into the relatively obvious gray-scale map of feature；
Gray-scale map is carried out the processing that labels by step 4, then, by the use of gray-scale map sum 70% as training set, uses remaining
30% is used as test set；
The gray-scale map of classification is converted to LDMB forms by step 5；
Step 6 by LDMB forms to gray-scale map be sent in depth network, tested by LeNet network trainings, analysis knot
Fruit.
2. a kind of ship noise method for identifying and classifying based on depth convolutional network according to claim one, its feature exist
In described MFCC processing, comprising the steps of：
Step 1 carries out preemphasis to the data of reading first, by a high-pass filter, and the purpose of preemphasis is to difference
Frequency range adds weights, improves HFS, makes the frequency spectrum of signal become flat；Formula is as follows：
H (Z)=1- μ Z-1 (1)
μ value is between 0.9-1.0 in formula 1；
N number of sampling point set is first synthesized an observation unit, referred to as frame by step 2；N value is 256 or 512 under normal circumstances,
The time covered is about 20~30ms or so；In order to avoid the change of adjacent two frame is excessive, allowing between two consecutive frames has one section of weight
Folded region, this overlapping region contain M sample point, and M value is about the 1/2 or 1/3 of N；
Step 3 uses Hamming window, to increase the continuity on the frame left side and the right in this stage of adding window；
Signal is transformed from the time domain to frequency domain by step 4 by fast Fourier；
Triangle filtering group of the step 5 by energy spectrum by one group of MEL yardstick, the number of triangular filter typically choose 18-22
It is individual optimal；
Step 6 obtains MFCC coefficient by discrete cosine transform.
3. a kind of ship noise method for identifying and classifying based on depth convolutional network according to claim one, its feature exist
In described LeNet networks, mainly comprising convolutional layer, lower sampling layer, connecting layer entirely；The convolution for the 5*5 sizes that convolutional layer uses
Core, and convolution kernel slides a pixel every time, a characteristic spectrum uses same convolution kernel, the ginseng in the value of each upper layer node
Numerical value, these products and an offset parameter are added to obtain an output, this is input to an activation primitive, swashed
The output of function living is the value of next node；Lower sampling layer is using 2*2 input domain, 4 nodes work of last layer
For the input of next layer of a node, and input domain is not overlapping, i.e., slides 2 pixels every time；4 of each lower sampling node are defeated
It is averaged after ingress summation, average is multiplied by a parameter and adds input of the offset parameter as an activation primitive, swashs
The output of function living is the input of next node layer；Sampling is only with two training parameters under one；Full articulamentum, i.e., a certain layer
Some node and each node of last layer connect, and each node each uses set of parameter, in the network connected entirely,
If K layers have n node, K+1 layers have m node；A then shared n*m connection, has one partially plus each K+1 node layers outside
Put, then share n*m+m training parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710717199.3A CN107609488B (en) | 2017-08-21 | 2017-08-21 | Ship noise identification and classification method based on deep convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710717199.3A CN107609488B (en) | 2017-08-21 | 2017-08-21 | Ship noise identification and classification method based on deep convolutional network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107609488A true CN107609488A (en) | 2018-01-19 |
| CN107609488B CN107609488B (en) | 2020-11-03 |
Family
ID=61064991
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710717199.3A Expired - Fee Related CN107609488B (en) | 2017-08-21 | 2017-08-21 | Ship noise identification and classification method based on deep convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107609488B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108847244A (en) * | 2018-08-22 | 2018-11-20 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Voiceprint recognition method and system based on MFCC and improved BP neural network |
| CN109100710A (en) * | 2018-06-26 | 2018-12-28 | 东南大学 | A kind of Underwater targets recognition based on convolutional neural networks |
| CN109547374A (en) * | 2018-11-23 | 2019-03-29 | 泰山学院 | A kind of depth residual error network and system for subsurface communication Modulation Identification |
| CN109767785A (en) * | 2019-03-06 | 2019-05-17 | 河北工业大学 | Ambient noise method for identifying and classifying based on convolutional neural networks |
| CN110738115A (en) * | 2019-09-12 | 2020-01-31 | 浙江大学 | propeller cavitation degree identification method based on pulse frequency characteristic pattern identification |
| CN110827844A (en) * | 2019-10-10 | 2020-02-21 | 天津大学 | Noise classification method based on BP network |
| CN110933235A (en) * | 2019-11-06 | 2020-03-27 | 杭州哲信信息技术有限公司 | Noise removing method in intelligent calling system based on machine learning |
| CN110956197A (en) * | 2019-10-28 | 2020-04-03 | 新绎健康科技有限公司 | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network |
| CN111986683A (en) * | 2020-09-04 | 2020-11-24 | 国家深海基地管理中心 | Method and system for evaluating deep sea ecosystem by using bioacoustic characteristics |
| CN112133941A (en) * | 2020-10-10 | 2020-12-25 | 西南交通大学 | Rapid fault diagnosis method for locomotive proton exchange membrane fuel cell system |
| CN112230205A (en) * | 2020-10-16 | 2021-01-15 | 哈尔滨工程大学 | Underwater target recognition system performance evaluation method using ship radiation noise simulation signal |
| CN113011093A (en) * | 2021-03-15 | 2021-06-22 | 哈尔滨工程大学 | Ship navigation noise simulation generation method based on LCWaveGAN |
| CN113012714A (en) * | 2021-02-22 | 2021-06-22 | 哈尔滨工程大学 | Acoustic event detection method based on pixel attention mechanism capsule network model |
| CN113739905A (en) * | 2020-05-27 | 2021-12-03 | 现代摩比斯株式会社 | Apparatus and method for locating noise occurring in steering system |
| CN114429771A (en) * | 2022-04-02 | 2022-05-03 | 武汉地震工程研究院有限公司 | Intelligent detection method and system for bonding defects of steel beam and CFRP (carbon fiber reinforced plastics) plate |
| CN114462457A (en) * | 2022-04-11 | 2022-05-10 | 自然资源部第一海洋研究所 | Ship underwater noise deep learning identification method based on intrinsic probability density function |
| GB2607290A (en) * | 2021-05-28 | 2022-12-07 | Bae Systems Plc | Apparatus and method of classification |
| CN117995193A (en) * | 2024-04-02 | 2024-05-07 | 山东天意装配式建筑装备研究院有限公司 | Intelligent robot voice interaction method based on natural language processing |
-
2017
- 2017-08-21 CN CN201710717199.3A patent/CN107609488B/en not_active Expired - Fee Related
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109100710A (en) * | 2018-06-26 | 2018-12-28 | 东南大学 | A kind of Underwater targets recognition based on convolutional neural networks |
| CN108847244A (en) * | 2018-08-22 | 2018-11-20 | 华东计算技术研究所（中国电子科技集团公司第三十二研究所） | Voiceprint recognition method and system based on MFCC and improved BP neural network |
| CN109547374A (en) * | 2018-11-23 | 2019-03-29 | 泰山学院 | A kind of depth residual error network and system for subsurface communication Modulation Identification |
| CN109767785A (en) * | 2019-03-06 | 2019-05-17 | 河北工业大学 | Ambient noise method for identifying and classifying based on convolutional neural networks |
| CN110738115A (en) * | 2019-09-12 | 2020-01-31 | 浙江大学 | propeller cavitation degree identification method based on pulse frequency characteristic pattern identification |
| CN110738115B (en) * | 2019-09-12 | 2022-11-15 | 浙江大学 | Propeller cavitation degree identification method based on pulse frequency characteristic pattern identification |
| CN110827844A (en) * | 2019-10-10 | 2020-02-21 | 天津大学 | Noise classification method based on BP network |
| CN110956197A (en) * | 2019-10-28 | 2020-04-03 | 新绎健康科技有限公司 | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network |
| CN110933235A (en) * | 2019-11-06 | 2020-03-27 | 杭州哲信信息技术有限公司 | Noise removing method in intelligent calling system based on machine learning |
| CN110933235B (en) * | 2019-11-06 | 2021-07-27 | 杭州哲信信息技术有限公司 | Noise identification method in intelligent calling system based on machine learning |
| CN113739905A (en) * | 2020-05-27 | 2021-12-03 | 现代摩比斯株式会社 | Apparatus and method for locating noise occurring in steering system |
| CN111986683A (en) * | 2020-09-04 | 2020-11-24 | 国家深海基地管理中心 | Method and system for evaluating deep sea ecosystem by using bioacoustic characteristics |
| CN112133941A (en) * | 2020-10-10 | 2020-12-25 | 西南交通大学 | Rapid fault diagnosis method for locomotive proton exchange membrane fuel cell system |
| CN112230205A (en) * | 2020-10-16 | 2021-01-15 | 哈尔滨工程大学 | Underwater target recognition system performance evaluation method using ship radiation noise simulation signal |
| CN113012714A (en) * | 2021-02-22 | 2021-06-22 | 哈尔滨工程大学 | Acoustic event detection method based on pixel attention mechanism capsule network model |
| CN113012714B (en) * | 2021-02-22 | 2022-08-02 | 哈尔滨工程大学 | Acoustic event detection method based on pixel attention mechanism capsule network model |
| CN113011093A (en) * | 2021-03-15 | 2021-06-22 | 哈尔滨工程大学 | Ship navigation noise simulation generation method based on LCWaveGAN |
| GB2607290A (en) * | 2021-05-28 | 2022-12-07 | Bae Systems Plc | Apparatus and method of classification |
| CN114429771A (en) * | 2022-04-02 | 2022-05-03 | 武汉地震工程研究院有限公司 | Intelligent detection method and system for bonding defects of steel beam and CFRP (carbon fiber reinforced plastics) plate |
| CN114462457A (en) * | 2022-04-11 | 2022-05-10 | 自然资源部第一海洋研究所 | Ship underwater noise deep learning identification method based on intrinsic probability density function |
| CN117995193A (en) * | 2024-04-02 | 2024-05-07 | 山东天意装配式建筑装备研究院有限公司 | Intelligent robot voice interaction method based on natural language processing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107609488B (en) | 2020-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107609488A (en) | A kind of ship noise method for identifying and classifying based on depth convolutional network | |
| Liu et al. | Underwater target recognition using convolutional recurrent neural networks with 3-D Mel-spectrogram and data augmentation | |
| CN101303764B (en) | Method for self-adaption amalgamation of multi-sensor image based on non-lower sampling profile wave | |
| CN107393542A (en) | A kind of birds species identification method based on binary channels neutral net | |
| CN106909784A (en) | Epileptic electroencephalogram (eeg) recognition methods based on two-dimentional time-frequency image depth convolutional neural networks | |
| CN110428842A (en) | Speech model training method, device, equipment and computer readable storage medium | |
| González-Hernández et al. | Marine mammal sound classification based on a parallel recognition model and octave analysis | |
| CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
| CN108846323A (en) | A kind of convolutional neural networks optimization method towards Underwater Targets Recognition | |
| Mishachandar et al. | Diverse ocean noise classification using deep learning | |
| CN108922513A (en) | Speech differentiation method, apparatus, computer equipment and storage medium | |
| CN104732970B (en) | A kind of ship-radiated noise recognition methods based on comprehensive characteristics | |
| CN109036460A (en) | Method of speech processing and device based on multi-model neural network | |
| Wei et al. | A method of underwater acoustic signal classification based on deep neural network | |
| Sun et al. | Underwater single-channel acoustic signal multitarget recognition using convolutional neural networks | |
| CN103208097A (en) | Principal component analysis collaborative filtering method for image multi-direction morphological structure grouping | |
| CN108922514A (en) | A kind of robust features extracting method based on low frequency logarithmic spectrum | |
| CN107368840A (en) | A kind of underwater manoeuvre Small object recognition methods based on MFCC and SVMs | |
| White et al. | More than a whistle: Automated detection of marine sound sources with a convolutional neural network | |
| CN112307926B (en) | Acoustic passive ship target classification method based on generation countermeasure network | |
| CN113642245B (en) | Construction method of ship radiation noise data set | |
| CN110211569A (en) | Real-time gender identification method based on voice map and deep learning | |
| Farrokhrooz et al. | Ship noise classification using probabilistic neural network and AR model coefficients | |
| Yi et al. | Multiresolution convolutional neural network for underwater acoustic target recognition | |
| CN112908344A (en) | Intelligent recognition method, device, equipment and medium for bird song |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201103 Termination date: 20210821 |