CN116910448A - Noise reduction method for weak speed signal of laser shock wave - Google Patents
Noise reduction method for weak speed signal of laser shock wave Download PDFInfo
- Publication number
- CN116910448A CN116910448A CN202310849718.7A CN202310849718A CN116910448A CN 116910448 A CN116910448 A CN 116910448A CN 202310849718 A CN202310849718 A CN 202310849718A CN 116910448 A CN116910448 A CN 116910448A
- Authority
- CN
- China
- Prior art keywords
- network model
- signal
- layer
- noise reduction
- noise
- 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
- 238000000034 method Methods 0.000 title claims abstract description 64
- 230000009467 reduction Effects 0.000 title claims abstract description 50
- 230000035939 shock Effects 0.000 title claims abstract description 35
- 239000002245 particle Substances 0.000 claims abstract description 60
- 238000012549 training Methods 0.000 claims abstract description 48
- 238000003062 neural network model Methods 0.000 claims abstract description 34
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 4
- 238000012512 characterization method Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 17
- 230000007246 mechanism Effects 0.000 claims description 17
- 238000012360 testing method Methods 0.000 claims description 13
- 238000004088 simulation Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 10
- 238000005259 measurement Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 238000001914 filtration Methods 0.000 abstract description 13
- 239000010410 layer Substances 0.000 description 67
- 238000010586 diagram Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 7
- 229910001069 Ti alloy Inorganic materials 0.000 description 6
- 239000002131 composite material Substances 0.000 description 6
- 239000011159 matrix material Substances 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 239000000126 substance Substances 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 239000012790 adhesive layer Substances 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000002238 attenuated effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000005474 detonation Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000002309 gasification Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000033001 locomotion Effects 0.000 description 1
- 239000007769 metal material Substances 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 239000013077 target material Substances 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/25—Design optimisation, verification or simulation using particle-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Software Systems (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
The invention discloses a noise reduction method of a weak speed signal of laser shock waves, which comprises the following steps: collecting particle velocity sample signals; constructing an attention-guided multi-scale convolutional neural network model; training the neural network model by using a particle velocity sample signal to obtain a trained network model; and inputting the particle speed signal to be detected into a trained network model, and outputting the particle speed signal after noise reduction. The invention can quickly and efficiently perform noise reduction treatment on the weak signals of different types of laser shock waves without relying on manual setting of filtering parameters, and effectively remove blind source noise in the weak echo signals.
Description
Technical Field
The invention relates to the field of laser shock wave interface bonding strength detection, in particular to a noise reduction method for a weak speed signal of laser shock waves.
Background
At high power (GW/cm) 2 Above) short pulse (ns below) laser and substance interaction process, the substance absorbs laser energy rapidly, gasification and plasma explosion occur, high pressure shock wave with pressure up to GPa and even TPa scale is formed, which makes laser induced shock wave applicable such as: the method is applied to the fields of laser shock reinforcement, laser shock wave interface bonding strength detection and the like.
Laser shock wave characteristics can be studied by measuring the free surface particle velocity of a target, and in recent years, measurement by using a Photon Doppler Velocimeter (PDV) in shock waves, detonation waves and other short-time high-speed motions is widely used. However, in PDV measurement, when the original signal is preprocessed, due to the low initial speed, the signal frequency and the low noise frequency are overlapped, so that not only the low frequency noise but also the low speed signal in the initial stage can be filtered, but also the noise can affect the speed extraction in the initial stage and even the speed extraction in other periods without filtering, so that a reasonable noise reduction filtering method becomes a key.
At present, the existing noise reduction filtering method can be divided into 8 methods such as an arithmetic average method, a polynomial fitting method, a median filtering method, a standard deviation filtering method, a classical Butterworth filter, a laser interference signal time-frequency domain filtering method, continuous wavelet transformation, fast Fourier transformation, multi-resolution singular value decomposition and the like, but the first 5 methods have the defects of starting from global filtering, cannot achieve high-frequency and low-frequency simultaneous consideration, and the last 3 time-frequency transformation methods have the problem of parameter optimization, the filtering effect is directly determined by parameter setting, and experimental results are greatly influenced by artificial experience.
In summary, it is difficult to retain useful echo signals of the shockwave from microvolts to millivolts on the back of the material on the premise of removing the interference of background noise by the existing noise reduction filtering method. Therefore, there is a need for a method of noise reduction of laser shock wave weak velocity signals that can solve the above problems.
Disclosure of Invention
Therefore, the invention aims to overcome the defects in the prior art, and provides the noise reduction method for the weak speed signal of the laser shock wave, which can quickly and efficiently perform noise reduction treatment on the weak signals of the laser shock wave of different types without depending on manual setting of filtering parameters, and effectively remove blind source noise in the weak echo signals.
The invention relates to a noise reduction method of a laser shock wave weak speed signal, which comprises the following steps:
collecting particle velocity sample signals;
constructing an attention-guided multi-scale convolutional neural network model;
training the neural network model by using a particle velocity sample signal to obtain a trained network model;
and inputting the particle speed signal to be detected into a trained network model, and outputting the particle speed signal after noise reduction.
Further, the particle velocity sample signal includes measured data and simulation data.
Further, the neural network model comprises a multi-scale sparse feature characterization module, a channel attention mechanism module, a feature enhancement module, an attention mechanism guiding module and a feature reconstruction module;
the multi-scale sparse feature characterization module is used for carrying out multi-scale feature extraction on the particle velocity signals to obtain multi-scale feature information;
the channel attention mechanism module is used for carrying out granularity processing on the multi-scale characteristic information to obtain different fine granularity characterization information of the noise-containing characteristic;
the feature enhancement module is used for carrying out information fusion on the characterization information with different fine granularity to obtain fused feature information;
the attention mechanism guiding module is used for adjusting the weights of the fused characteristic information at different spatial positions and screening out characteristic signals representing noise;
the characteristic reconstruction module is used for differentiating the original particle velocity signal and the characteristic signal representing noise to obtain a particle velocity signal after noise reduction.
Further, training the neural network model by using the particle velocity sample signal to obtain a trained network model, which specifically comprises:
dividing the particle velocity sample signal into a training set and a test set;
setting training rounds as K times, and determining a damage function; the damage function is the difference value between the particle velocity signal after noise reduction and the actual pure particle velocity signal;
iterative training of the neural network model using a training set, wherein in the iterative training process, the first k 1 The learning rate in the iterative training is set to be tau 1 The method comprises the steps of carrying out a first treatment on the surface of the Kth 2 Secondary to k 3 The learning rate in the iterative training is set to tau 2 The method comprises the steps of carrying out a first treatment on the surface of the Last k 4 The learning rate in the iterative training is set to tau 3 The method comprises the steps of carrying out a first treatment on the surface of the Wherein τ 1 >τ 2 >τ 3 ;
Judging whether the loss function meets the target requirement, if so, finishing training, and taking the network model at the moment as a trained network model; if not, the network model parameters are adjusted, iteration training is continued until the iteration times are K times, and the network model after the last training is used as a trained network model.
Further, in the iterative training process of the neural network model, the fine tuning data set is used for fine tuning the neural network model, which specifically comprises the following steps:
taking a multi-scale convolution layer in the neural network model as a target layer, and determining the layer number U of the target layer;
inputting the fine tuning data set into the neural network model, locking a first layer in the target layer, adjusting model parameters of other layers after the first layer is removed in the target layer, so that the neural network model meets the first fine tuning requirement, locking a second layer in the target layer, adjusting model parameters of other layers after the first layer and the second layer are removed in the target layer, so that the neural network model meets the second fine tuning requirement, and so on, locking other layers in the target layer in sequence, and adjusting model parameters of a layer which is not locked in the target layer until a U-1 layer in the target layer is locked, and adjusting model parameters of a U layer in the target layer.
Further, a trim dataset is made according to the following method:
a. setting initial working condition parameters under the set size of the light spot of the shock wave, wherein the initial working condition parameters comprise an initial pulse width sigma 0 Initial shock wave energy E 0 ;
b. Under the initial working condition, the initial pulse width sigma is maintained 0 Unchanged at energy E 0 On the basis of increasing the energy value each time, carrying out multiple collection to obtain a plurality of pieces of particle speed signal data;
c. on the basis of the initial working condition parameters, the pulse width sigma is obtained after the pulse width is increased 1 The energy E is obtained after the energy of the impact wave is increased 1 Forming a second working condition parameter;
d. in the second working condition, the pulse width sigma is maintained 1 Unchanged at energy E 1 On the basis of increasing the energy value each time, carrying out multiple collection to obtain a plurality of pieces of particle speed signal data;
e. and c, analogizing according to the steps c-d, acquiring particle speed signal data under other working conditions, and taking the particle speed signal data acquired under all the working conditions as a fine adjustment data set.
Further, the simulation data are simulation data added with white noise with different signal to noise ratios.
Further, a particle velocity sample signal is acquired using a detector and a voltage V (t) of the particle velocity sample signal is determined according to the following formula:
wherein R is the voltage sensitivity of the detector, I 1 、I 2 The reference light intensity and the sensing light intensity of the detector are respectively, cos [. Cndot.]Representing a cosine function, f beat (t) represents the difference frequency as a function of the measurement time t,is the difference between the primary phases of the reference light and the sensing light.
Further, the multi-scale sparse feature characterization module takes a common convolution layer as a low-energy feature characterization layer and takes a multi-scale convolution layer as a high-energy feature characterization layer.
Further, testing the trained network model by using a test set, and evaluating a test result by adopting a signal-to-noise ratio; the signal-to-noise ratio SNR is determined according to the following formula:
wherein V is s Representing the voltage value of the noisy signal, V n Representing the voltage value of the noise signal.
The beneficial effects of the invention are as follows: according to the noise reduction method for the laser shock wave weak speed signal, disclosed by the invention, different fine granularity characterization of the noise-containing signal characteristic is realized by establishing a multi-scale sparse characteristic extraction network and combining with ECA channel module self-adaption; the deep layer characteristics are fully learned by utilizing jump connection to fuse the information of each channel of the deep layer characteristics/shallow layer characteristics of the network, the influence of the shallow layer characteristics is not ignored, finally, the weights of different positions of the characteristic space are adjusted through an attention mechanism, the accurate representation of random noise is obtained, the learned noise characteristics are removed in noise-containing data through a reconstruction module, and therefore the accurate noise reduction of weak signals is achieved. The method is independent of manual experience setting parameters, effectively removes blind source noise in the weak echo signals, and has the technical characteristics of simplicity, easiness in operation, accuracy and high efficiency.
Drawings
The invention is further described below with reference to the accompanying drawings and examples:
FIG. 1 is a schematic diagram of a noise reduction method of the present invention;
FIG. 2 is a diagram of a neural network model network of the present invention;
FIG. 3 is a schematic diagram showing the simulation data acquisition effect of the present invention;
FIG. 4 is a schematic diagram showing the comparison of the noise reduction method of the present invention and the conventional noise reduction method;
FIG. 5 is a graph comparing original data of the titanium alloy with noise reduction effect of the present invention;
FIG. 6 is a schematic diagram of a composite raw data particle velocity signal according to the present invention;
FIG. 7 is a schematic diagram of particle velocity signals after noise reduction of the composite material of the present invention.
Detailed Description
The invention is further described with reference to the accompanying drawings, in which:
the invention relates to a noise reduction method for a laser shock wave weak speed signal, which comprises the following steps:
s1, collecting particle velocity sample signals;
s2, constructing an attention-guided multi-scale convolutional neural network model;
s3, training the neural network model by using a particle velocity sample signal to obtain a trained network model;
s4, inputting the particle speed signal to be detected into the trained network model, and outputting the particle speed signal after noise reduction.
The method obtains the distributed characteristic expression of the noise signals, guides the noise-containing signals to train by taking the preprocessed signals as labels, finally obtains the accurate representation of the noise signals in a mode that the noise-containing signals and the quasi-pure signals are poor, and solves the problem that the traditional global filtering method cannot set reasonable starting and cut-off frequencies to cause excessive filtering when the weak signal noise signals are mixed with the useful signals.
In this embodiment, the particle velocity sample signal includes measured data and simulation data. The simulation data are simulation data added with white noise with different signal to noise ratios.
The method comprises the steps of obtaining pure speed signals according to simulation of experimentally obtained laser impact target back particle speed signals, wherein the back particle speed signals are obtained through continuous wavelet transformation as shown in a back particle speed part in fig. 3; after the impact wave acts on the target, the impact wave can travel back and forth on the front surface and the rear surface of the target and is attenuated with energy, so that the particle speed of the target has similar periodicity, and on the basis of a particle speed signal, a voltage signal V (t) measured by a PDV (photon Doppler velocimeter) is obtained by back-pushing according to a formula:
wherein: r is the voltage sensitivity of the detector, I 1 、I 2 The reference light intensity and the sensing light intensity of the detector are respectively, cos [. Cndot.]Representing a cosine function, f beat (t) represents the difference frequency as a function of the measurement time t,is the difference between the primary phases of the reference light and the sensing light. The voltage signal measured by the photon doppler velocimeter is an electrical signal converted from scattered light and contains information about the frequency change due to the doppler effect. By analyzing and processing this signal, it can be countedAnd calculating the speed of the target object.
To simulate the actual situation, white noise (used for simulating electronic noise generated by an oscilloscope and a detector) and low-frequency noise (used for simulating low-frequency noise generated by the change of the reflection angle of the target material) are added into the pushed voltage signal, and the range is set to be about 25 mv. The left area of the dashed box in the amplitude part of fig. 3 is the electrical noise of the PDV detector before being impacted, and is also considered, and finally the left area of the dashed box is an analog signal, so that only 2us is shown in the figure for convenience of distinguishing.
Because the sampling rate is selected to be 10GS/s in the acquisition process of the oscilloscope for actually measured data, and the acquisition time length is 20us, the total measurement is 2.0x10 5 And selecting 800 simulation curves for adding different signal to noise ratio white noise (1 step in the range of 10-40) respectively. And combining 1000 pieces of actual data of 200 pieces of actual data, wherein 600 pieces of actual data are used for training, 400 pieces of actual data are used for testing, and the training set and the testing set are randomly sampled and divided. The input matrix size is 200000x1000 in matrix form.
In this embodiment, in step S2, as shown in fig. 2, the neural network model includes a multi-scale sparse feature characterization module, a channel attention mechanism module, a feature enhancement module, an attention mechanism guiding module, and a feature reconstruction module; wherein, the channel attention mechanism can adopt an ECA channel attention mechanism;
the multi-scale sparse feature characterization module is used for carrying out multi-scale feature extraction on the particle velocity signals to obtain multi-scale feature information;
the multi-scale sparse feature characterization module takes a common convolution layer as a low-energy feature characterization layer and takes a multi-scale convolution layer as a high-energy feature characterization layer; the combination representation of a few high-energy points and a plurality of low-energy points is regarded as a novel feature sparse mechanism, and the mechanism can effectively improve the noise reduction performance and efficiency; the common convolution layer is a type of neural network layer commonly used in deep learning, and is mainly used for processing images and other similar structured data.
The channel attention mechanism module is used for carrying out granularity processing on the multi-scale characteristic information to obtain different fine granularity characterization information of the noise-containing characteristic; the fine granularity is only determined by the size of the receptive field of the multi-scale convolution network, the ECA can dynamically and adaptively adjust the size of the convolution kernel to capture information among different channels and interact, so that the channel dimension is reduced and the input parameter quantity is reduced when the channel attention information is learned.
The feature enhancement module is used for carrying out information fusion on the characterization information with different fine granularity to obtain fused feature information; the feature enhancement module effectively fuses deep/shallow information by utilizing jump connection, adds element levels of feature layers of different layers, and fully learns deep features without neglecting shallow feature influence.
The attention mechanism guiding module is used for adjusting the weights of the fused characteristic information at different spatial positions and screening out characteristic signals representing noise;
the characteristic reconstruction module is used for differentiating the original particle velocity signal and the characteristic signal representing noise to obtain a particle velocity signal after noise reduction.
In this embodiment, in step S3, the neural network model is trained using the particle velocity sample signal to obtain a trained network model, which specifically includes:
dividing the particle velocity sample signal into a training set and a test set; because the sample signal data is a time sequence, the total is 2.0x10 5 As a batch input; wherein, the 0.5mm thick titanium alloy impact signal can be used as an embodiment, the signal strength can be effectively ensured because the titanium alloy impact signal is an isotropic metal material, and the propagation rule of the impact wave in the titanium alloy impact signal is obvious;
setting training rounds as 100 times, and determining a damage function; the damage function is the difference value between the particle velocity signal after noise reduction and the actual pure particle velocity signal; the existing ADAM algorithm can be adopted to optimize the loss function;
performing iterative training on the neural network model by using a training set, wherein the learning rate in the first 50 iterative training is set to be 0.01 in the iterative training process; the learning rate in the iterative training from 60 th to 90 th is set to 0.001; the learning rate in the last 10 iterative training is set to 0.0001;
judging whether the loss function meets the target requirement, if so, finishing training, saving model parameters, and taking the network model at the moment as a trained network model; if not, the network model parameters are adjusted, iteration training is continued until the iteration times are 100 times, and the network model after the last training is used as a trained network model. The target requirement can be set or determined according to the actual working condition.
In this embodiment, in order to increase the migration performance of the network, a fine tuning manner is used to enable the neural network model to adapt to different working conditions and different types of data;
in the iterative training process of the neural network model, the fine tuning data set is used for fine tuning the neural network model, and the method specifically comprises the following steps:
taking a multi-scale convolution layer in the neural network model as a target layer, and determining the layer number U of the target layer; the multi-scale sparse feature characterization module network structure parameters are shown in table 1, wherein a multi-scale convolution layer in the multi-scale sparse feature characterization module is used as a target layer:
TABLE 1
In table 1, input represents an Input; PSConv represents a multi-scale convolutional layer, with u=4 layers in total; the enhancement layer can enhance the characteristics so as to improve the fitting capacity and robustness of the network model, and BN represents batch normalization.
And (3) inputting a fine tuning data set into the neural network model, locking a first layer in the target layer, adjusting model parameters of other layers after the first layer is removed in the target layer, so that the neural network model meets the first fine tuning requirement, locking a second layer in the target layer, adjusting model parameters of other layers after the first layer and the second layer are removed in the target layer, so that the neural network model meets the second fine tuning requirement, and so on, locking other layers in the target layer in sequence, and adjusting model parameters of a hierarchy which is not locked in the target layer until a 3 rd layer in the target layer is locked, and adjusting model parameters of a 4 th layer in the target layer. Wherein, a plurality of fine tuning requirements can be set or determined according to actual working conditions.
In this embodiment, the fine tuning dataset is made according to the following method:
setting initial working condition parameters under the condition that the diameter of a set shock wave spot is 5mm, wherein the initial working condition parameters comprise an initial pulse width of 20ns and initial shock wave energy of 2-3J;
under the initial working condition, the initial pulse width is kept unchanged for 20ns, the energy is stepped for 0.2J each time for 2-3J, and 48 data can be obtained;
setting a second working condition parameter: pulse width is 100ns, spot diameter is 5mm, energy is 2-16J step by step for 2J each time, and 32 data can be obtained;
similarly, the pulse width is 200ns, the diameter of a light spot is 5mm, the energy is 4-32J, and each time the energy is stepped by 4J, 32 data can be obtained; pulse width 300ns, spot diameter 5mm, energy 6-48J step by step 6J each time, can get the data 32;
further, the same position was partially impacted 2 times, and 56 pieces of data such as measurement were repeated. The power density is kept in the range of 0.1-0.8 GW.cm -2 A total of 200. Each sampling frequency was 10GS/s, the acquisition time length was 20us, and a total of 2.0x10 were measured 5 And sampling points. A total of 200 pieces are used as fine tuning data sets, and the matrix size is 200000x200 in matrix form. Wherein all data is input with the original data as input, and the preprocessed data as tag.
Taking the impact signal of the titanium alloy with the thickness of 0.5mm as an example, the effect comparison chart shown in fig. 5 can be obtained after the noise reduction method of the invention. In addition, by comparing the noise reduction method of the present invention with the conventional noise reduction method, a comparison diagram as shown in fig. 4 can be obtained, wherein the noise reduction method of the present invention is denoted by reference numeral 7 in fig. 4.
In order to better understand the noise reduction method of the invention, a 3mm composite material laminated plate/cementing piece is used as a test material, the shock wave is required to face the conditions of multi-interface propagation coupling and reflection in the internal propagation process of the material, and a thin adhesive layer is used for blocking in the cementing process, so that the signal is weaker.
In order to prove the advantage of the noise reduction method for weak signals, the neural network model is finely adjusted by utilizing impact data of the 3mm composite material laminated plate/cementing piece, so that the intrinsic characteristics of the composite material impulse signals are learned.
Specifically, different working condition data are selected, wherein the diameter of a light spot is 5mm, and other parameters are as follows: pulse width 20ns, energy 2-4J steps 0.4J data 24 pieces each time; pulse width 50ns, energy 3-6J steps 1J data 16 pieces each time; pulse width is 100ns, energy 1-7J steps 1J data 28 pieces each time; pulse width is 200ns, energy is 1-9J, and 1J data 36 pieces are stepped each time; pulse width 300ns, energy 1-10J steps 1J data 40 pieces each time; 144 pieces of relevant data such as 36 parameters, 4 points of impact of each parameter, and 200 pieces of relevant data such as repeated test are combined, and the power density is in the range of 0.05-0.51 GW.cm -2 The signal intensity is far less than that of titanium alloy experiments.
With composite material data 200000x200 as a fine tuning data set, the experimental iteration is performed 100 times, the learning rate in the first 50 training iterations of the learning rate is set to 0.01, the learning rate in the period from 60 to 90 training iterations is set to 0.001, the last 10 learning rates are set to 0.0001 to dynamically change, the sizes of 1-4 layers of multi-scale convolution kernels are 7, 5, 3 and 3 in sequence, and the ECA channel number is set to 64.
Defining a loss function by taking the difference value between the data after noise reduction and the actual pure signal as a target, performing model training and fine tuning, and storing parameters after training is finished;
randomly extracting 200 pieces of data from 400 pieces of data to be tested; the signal-to-noise ratio (Signal to Noise Ratio, SNR) is adopted as a noise reduction effect evaluation index, and the calculation formula is shown as the following formula:
wherein V is s Representing the voltage value of the noisy signal, V n Representing the voltage value of the noise signal. A higher signal-to-noise ratio corresponds to a stronger signal and less noise, and thus a higher signal-to-noise ratio is generally considered better.
After the signal-to-noise ratio index converges and keeps stable, noise reduction data of the final weak signal is obtained, and the effect is shown in fig. 6 and 7.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the technical solution of the present invention, which is intended to be covered by the scope of the claims of the present invention.
Claims (10)
1. A noise reduction method for a weak speed signal of laser shock waves is characterized by comprising the following steps of: comprising the following steps:
collecting particle velocity sample signals;
constructing an attention-guided multi-scale convolutional neural network model;
training the neural network model by using a particle velocity sample signal to obtain a trained network model;
and inputting the particle speed signal to be detected into a trained network model, and outputting the particle speed signal after noise reduction.
2. The method for noise reduction of laser shock wave weak velocity signals according to claim 1, wherein: the particle velocity sample signal includes measured data and simulation data.
3. The method for noise reduction of laser shock wave weak velocity signals according to claim 1, wherein: the neural network model comprises a multi-scale sparse feature characterization module, a channel attention mechanism module, a feature enhancement module, an attention mechanism guiding module and a feature reconstruction module;
the multi-scale sparse feature characterization module is used for carrying out multi-scale feature extraction on the particle velocity signals to obtain multi-scale feature information;
the channel attention mechanism module is used for carrying out granularity processing on the multi-scale characteristic information to obtain different fine granularity characterization information of the noise-containing characteristic;
the feature enhancement module is used for carrying out information fusion on the characterization information with different fine granularity to obtain fused feature information;
the attention mechanism guiding module is used for adjusting the weights of the fused characteristic information at different spatial positions and screening out characteristic signals representing noise;
the characteristic reconstruction module is used for differentiating the original particle velocity signal and the characteristic signal representing noise to obtain a particle velocity signal after noise reduction.
4. The method for noise reduction of laser shock wave weak velocity signals according to claim 1, wherein: training the neural network model by using a particle velocity sample signal to obtain a trained network model, which specifically comprises the following steps:
dividing the particle velocity sample signal into a training set and a test set;
setting training rounds as K times, and determining a damage function; the damage function is the difference value between the particle velocity signal after noise reduction and the actual pure particle velocity signal;
iterative training of the neural network model using a training set, wherein in the iterative training process, the first k 1 The learning rate in the iterative training is set to be tau 1 The method comprises the steps of carrying out a first treatment on the surface of the Kth 2 Secondary to k 3 The learning rate in the iterative training is set to tau 2 The method comprises the steps of carrying out a first treatment on the surface of the Last k 4 The learning rate in the iterative training is set to tau 3 The method comprises the steps of carrying out a first treatment on the surface of the Wherein τ 1 >τ 2 >τ 3 ;
Judging whether the loss function meets the target requirement, if so, finishing training, and taking the network model at the moment as a trained network model; if not, the network model parameters are adjusted, iteration training is continued until the iteration times are K times, and the network model after the last training is used as a trained network model.
5. The method for noise reduction of laser shock wave weak velocity signals according to claim 4, wherein: in the iterative training process of the neural network model, the fine tuning data set is used for fine tuning the neural network model, and the method specifically comprises the following steps:
taking a multi-scale convolution layer in the neural network model as a target layer, and determining the layer number U of the target layer;
inputting the fine tuning data set into the neural network model, locking a first layer in the target layer, adjusting model parameters of other layers after the first layer is removed in the target layer, so that the neural network model meets the first fine tuning requirement, locking a second layer in the target layer, adjusting model parameters of other layers after the first layer and the second layer are removed in the target layer, so that the neural network model meets the second fine tuning requirement, and so on, locking other layers in the target layer in sequence, and adjusting model parameters of a layer which is not locked in the target layer until a U-1 layer in the target layer is locked, and adjusting model parameters of a U layer in the target layer.
6. The method for noise reduction of laser shock wave weak velocity signals according to claim 5, wherein: the fine tuning dataset was made according to the following method:
a. setting initial working condition parameters under the set size of the light spot of the shock wave, wherein the initial working condition parameters comprise an initial pulse width sigma 0 Initial shock wave energy E 0 ;
b. Under the initial working condition, the initial pulse width sigma is maintained 0 Unchanged at energy E 0 On the basis of increasing the energy value each time, carrying out multiple collection to obtain a plurality of pieces of particle speed signal data;
c. on the basis of the initial working condition parameters, the pulse width sigma is obtained after the pulse width is increased 1 The energy E is obtained after the energy of the impact wave is increased 1 Forming a second working condition parameter;
d. in the second working condition, the pulse width sigma is maintained 1 Unchanged at energy E 1 On the basis of increasing the energy value each time, carrying out multiple collection to obtain a plurality of pieces of particle speed signal data;
e. and c, analogizing according to the steps c-d, acquiring particle speed signal data under other working conditions, and taking the particle speed signal data acquired under all the working conditions as a fine adjustment data set.
7. The method for noise reduction of laser shock wave weak velocity signals according to claim 2, wherein: the simulation data are simulation data added with white noise with different signal to noise ratios.
8. The method for noise reduction of laser shock wave weak velocity signals according to claim 1, wherein: a particle velocity sample signal is acquired using a detector and a voltage V (t) of the particle velocity sample signal is determined according to the formula:
wherein R is the voltage sensitivity of the detector, I 1 、I 2 The reference light intensity and the sensing light intensity of the detector are respectively, cos [. Cndot.]Representing a cosine function, f beat (t) represents the difference frequency as a function of the measurement time t,is the difference between the primary phases of the reference light and the sensing light.
9. A method of noise reduction of a laser shock wave weak velocity signal according to claim 3, wherein: the multi-scale sparse feature characterization module takes a common convolution layer as a low-energy feature characterization layer and takes a multi-scale convolution layer as a high-energy feature characterization layer.
10. The method for noise reduction of laser shock wave weak velocity signals according to claim 4, wherein: testing the trained network model by using a test set, and evaluating a test result by adopting a signal-to-noise ratio; the signal-to-noise ratio SNR is determined according to the following formula:
wherein V is s Representing the voltage value of the noisy signal, V n Representing the voltage value of the noise signal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310849718.7A CN116910448A (en) | 2023-07-11 | 2023-07-11 | Noise reduction method for weak speed signal of laser shock wave |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310849718.7A CN116910448A (en) | 2023-07-11 | 2023-07-11 | Noise reduction method for weak speed signal of laser shock wave |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116910448A true CN116910448A (en) | 2023-10-20 |
Family
ID=88354282
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310849718.7A Pending CN116910448A (en) | 2023-07-11 | 2023-07-11 | Noise reduction method for weak speed signal of laser shock wave |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116910448A (en) |
-
2023
- 2023-07-11 CN CN202310849718.7A patent/CN116910448A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104914446B (en) | Three-dimensional distance image time domain real-time de-noising method based on photon counting | |
CN102436586B (en) | Hyper spectral image classification method based on wavelet threshold denoising and empirical mode decomposition | |
CN103995289B (en) | Time-varying method of mixed phase wavelet extraction based on time-frequency spectrum analog | |
CN110865357A (en) | Laser radar echo signal noise reduction method based on parameter optimization VMD | |
CN108267784A (en) | A kind of seismic signal random noise compression process method | |
CN113642484B (en) | Magnetotelluric signal noise suppression method and system based on BP neural network | |
CN108680955B (en) | A kind of Lithology Identification Methods and identifying system | |
CN114897023B (en) | Underwater sound target identification method based on extraction of sensitive difference features of underwater sound targets | |
CN115499092B (en) | Astronomical radio transient signal searching method, system, device and readable storage medium | |
CN103344989A (en) | Method for analyzing impulse noise interference in vibroseis earthquake records | |
Li et al. | Denoising seismic signal via resampling local applicability functions | |
Han et al. | A novel rolling bearing fault diagnosis method based on generalized nonlinear spectral sparsity | |
Wang et al. | Radar HRRP target recognition in frequency domain based on autoregressive model | |
CN107632326A (en) | Geophysical signal denoising method | |
CN109507292A (en) | A kind of method for extracting signal | |
CN116910448A (en) | Noise reduction method for weak speed signal of laser shock wave | |
Lin et al. | Research on microseismic denoising method based on CBDNet | |
CN111191502B (en) | Stick-slip and jump drill abnormal working condition identification method based on drill string vibration signal | |
CN110837088B (en) | Data denoising method for spaceborne laser altimeter | |
Basu et al. | A nonparametric test for stationarity based on local Fourier analysis | |
CN113221615A (en) | Partial discharge pulse extraction method based on noise reduction clustering | |
CN115563480A (en) | Gear fault identification method for screening octave geometric modal decomposition based on kurtosis ratio coefficient | |
CN111538086B (en) | First arrival automatic pickup method for improving seismic data first arrival wave quality | |
Jiang et al. | Automatic microseismic events detection using multiscale morphological characteristic function | |
CN108445541B (en) | It is concerned with the method for important body construction Steerable filter compacting acquisition footprint based on statistics |
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 |