CN104568730A - Electrochemical noise corrosion signal processing method based on neural network - Google Patents
Electrochemical noise corrosion signal processing method based on neural network Download PDFInfo
- Publication number
- CN104568730A CN104568730A CN201410847805.XA CN201410847805A CN104568730A CN 104568730 A CN104568730 A CN 104568730A CN 201410847805 A CN201410847805 A CN 201410847805A CN 104568730 A CN104568730 A CN 104568730A
- Authority
- CN
- China
- Prior art keywords
- corrosion
- neural network
- noise
- data
- training
- 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
Abstract
The invention discloses an electrochemical noise corrosion signal processing method based on a neural network. The electrochemical noise corrosion signal processing method comprises the following steps: 1) extracting typical electrochemical noise signals of different corrosion types; 2) performing data packet; 3) obtaining characteristic parameters of each group of data; 4) training and processing a BP (back propagation) neural network. Characteristic values of all file samples are calculated by using electrochemical noise corrosion signal data obtained through an electrochemical noise corrosion experiment that electrochemical noises are processed by the BP neural network; data processing results show that the neural network can distinguish the corrosion types at the precision as high as 98.3%, and the electrochemical noise corrosion signal processing method is an effective electrochemical noise data processing method.
Description
Technical field
The invention belongs to material corrosion detection technique field, particularly relate to a kind of electrochemistry noise based on neural network corrosion signal processing method.
Background technology
Electrochemistry noise disposal route novel in recent years emerges in an endless stream, and novel mathematical method is still used for reference in the process of electrochemistry noise data by its main thought, specifically has the following method: cluster analysis, divides the corrosion stage by the method for cluster; Chaology, distinguishes corrosion type by the parameter of chaos; Neural network, differentiates corrosion type and prediction etc. by the neural network of training; Although these new method new arguments are a lot, corresponding then slightly aobvious weak with corrosion process principle, needs the time more of a specified duration and more fully argument remove to prove validity and the practicality row of these methods; Therefore, how finding the disposal route of new electrochemistry noise data and it combined with its corrosion process principle, making this disposal route obtain rigorous theoretic foundation, is all the emphasis also will being the research of following electrochemistry noise in recent years.
Neural network (Neural Networks) belongs to the category of machine learning, and its basic structural unit is the node of neural network.A node is the elementary cell calculated, realize the arithmetic of the addition subtraction multiplication and division in calculating etc., countless nodes constitutes the neural network of whole calculating by the input and output of ad hoc structure, just as the human brain nervous system be made up of numerous neuronal, therefore it is in the nature mathematics computing model, be referred to as neural network visually.The purposes of neural network widely, is mainly used in pattern-recognition, data prediction, data fitting field.Neural network classification has a variety of method, can be divided into feedforward neural network and Feedback Neural Network by network topology structure, can be divided into supervised learning neural network and unsupervised learning neural network etc. by learning process.There is supervision type neural network can have better classification performance compared with unsupervised neural network, a large amount of Foundation is experiment inventions provides very abundant training sample, therefore adopt a kind of supervised learning neural network in this article----reverse transmittance nerve network, i.e. BP neural network.
Summary of the invention
In order to solve the problem, a kind of electrochemistry noise based on neural network is the object of the present invention is to provide to corrode signal processing method.
In order to achieve the above object, the corrosion of the electrochemistry noise based on neural network signal processing method provided by the invention comprises: the following step carried out in order:
Step 1) refine the typical electrochemical noise signal of different corrosion type:
Corrosion region is divided into spot corrosion, uniform corrosion and passivation three kinds of corrosion types by the time-domain diagram according to the electrochemical noise signal obtained from electrochemistry noise experiment; Described electrochemistry noise experiment makes working electrode with standard 304 stainless steel sample, and two working electrodes and a contrast electrode are put among experimental solutions, measure the electrochemical source of current noise between two working electrodes and the electrochemical voltage noise between working electrode and contrast electrode by zero resistance galvanometer simultaneously; The electrochemistry noise of its performance of different corrosion types has very large difference on time-domain diagram; When there is spot corrosion, current noise and voltage noise there will be obvious transient state peak; When there is passivation or uniform corrosion, current noise and voltage noise then show as the irregular vibration of high frequency;
Step 2) packet:
Be that a unit carries out data cutting by the above-mentioned experimental data collected by 1024 points, namely within every 512 seconds, deposit an independent small documents, filename temporally superposes, then obtain the data of a large amount of various corrosion;
Step 3) often organized each characteristic parameter of data:
To often kind of corrosion type random selecting up to a hundred groups of data in these all small documents, these data of batch processing, are often organized the value of each characteristic parameter of data;
Step 4) training of BP neural network and process:
By calculating the electrochemical signals eigenwert of each group, obtain the value of up to a hundred groups of samples; 10 parameters often organizing sample are formed a stack features vector by the order of the energy Ratios of noise resistance, feature electric charge, characteristic frequency, small echo normalized energy d1 ~ d7 layer, and up to a hundred groups of sample values can obtain up to a hundred groups of different proper vectors; Every stack features vector, as the input set of BP neural network, often organizes the output set of corrosion type corresponding to sample as BP neural network; 60% input of getting input total value, as training set, gets the input of 20% as checking collection and test set; Use training set training data, do the selection of neural metwork training optimal value with checking collection, by Generalization Capability and the nicety of grading of test set test neural network.
In step 4) in, need the optimization first carrying out BP neural network before the training of described BP neural network and process:
The optimizing process of BP neural network is the process selecting most suitable three parameters: hidden layer nodes, learning rate and learning function;
Select the most frequently used function
its Complete Classification training precision is high and generalization ability is strong;
The selection of hidden layer nodes with reference to experimental formula: l≤n-1 or
wherein a is the value between 0 ~ 10.
Electrochemistry noise based on neural network corrosion signal processing method provided by the invention is the electrochemistry noise corrosion signal data with BP Processing with Neural Network electrochemistry noise corrosion experiment gained, calculate the eigenwert of All Files sample, data processed result shows, neural network can pick out corrosion type with the precision up to 98.3%, is a kind of effective electrochemistry noise data processing method.
Accompanying drawing explanation
Fig. 1 is the process flow diagram of the electrochemistry noise based on neural network provided by the invention corrosion signal processing method;
Fig. 2 is the noise signal time-domain diagram of typical three kinds of corrosion types;
Fig. 3 is that sample characteristics enumerates figure;
Fig. 4 is BP neural metwork training performance map.
Embodiment
Below in conjunction with the drawings and specific embodiments, the corrosion of the electrochemistry noise based on neural network signal processing method provided by the invention is described in detail.
As shown in Figure 1, the corrosion of the electrochemistry noise based on neural network signal processing method provided by the invention comprises the following step performed in order:
Step 1) refine the typical electrochemical noise signal of different corrosion type:
Corrosion region is divided into spot corrosion, uniform corrosion and passivation three kinds of corrosion types by the time-domain diagram according to the electrochemical noise signal obtained from electrochemistry noise experiment; Described electrochemistry noise experiment makes working electrode with standard 304 stainless steel sample, and two working electrodes and a contrast electrode are put among experimental solutions, the electrochemical source of current noise between two working electrodes and the electrochemical voltage noise between working electrode and contrast electrode can be measured by zero resistance galvanometer (ZRA) simultaneously, obtain the typical electrochemical noise signal of different corrosion type thus; The electrochemistry noise of its performance of different corrosion types has very large difference on time-domain diagram; When there is spot corrosion, current noise and voltage noise there will be obvious transient state peak; When there is passivation or uniform corrosion, current noise and voltage noise then show as the irregular vibration of high frequency, and Fig. 2 is the time-domain diagram of three kinds of pink noise signal; Be respectively Fig. 2 a-Fig. 3 f from left to right, from top to bottom, wherein Fig. 2 a and Fig. 2 d shows the FeCl of standard 304 stainless steel sample at 0.2mol/l
3middlely soaked the current potential after 2 hours and current noise figure, can find that current potential noise has the transient state peak of peak value 3mV significantly, and current noise has peak value to be the transient state peak of 7 μ A, both transient state peaks occur at one time, but its direction is just in time contrary; Fig. 2 b and Fig. 2 e is then the H of sample at 0.6mol/l
2sO
4the current potential of middle immersion after 6 hours and current noise figure, its current noise and current potential noise show as the irregular vibration of high frequency, and its current potential amplitude is at 0.4mV, and current amplitude is at 200nA; Fig. 2 c and Fig. 2 f is current potential after sample soaks 2 hours in the mixed solution of 0.1mol/l NaOH and 0.1mol/l KOH and current noise figure, it shows the oscillator signal similar with uniform corrosion, but its current potential amplitude is only 0.5mV, and current amplitude is only 0.5nA, is significantly less than uniform corrosion;
In different solutions, soak the metallograph after 72 hours according to standard 304 stainless steel sample, original sample surface unusual light, the sample after spot corrosion shows specimen surface and occurs typical full grown some pit; Sample, after uniform corrosion, can see the austenite shape of standard 304 stainless steel sample at specimen surface; During sample passivation, there is passivating film in its surface, and be characterized as specimen surface by smooth roughening, color is tarnished by light; The metallograph of sample demonstrates the rationality of experimental design and the validity of experimental result well by visual characteristic method;
Step 2) packet:
Table 1 test for electrochemistry noise experimental solutions, corrosion type, duration and label summary sheet, the experimental data collected by experiment content according to table 1 is that a unit carries out data cutting by 1024 points, namely within every 512 seconds, an independent small documents is deposited, filename temporally superposes, then can obtain the data of a large amount of various corrosion;
Table 1 experimental solutions is summed up
Step 3) often organized each characteristic parameter of data:
To often kind of corrosion type random selecting 100 groups of data in these all small documents, these data of batch processing, are often organized the value of each characteristic parameter of data, and its result as shown in Figure 3; In figure 3, be respectively Fig. 3 a-Fig. 3 k from left to right, from top to bottom, front 100 groups of samples are tested in PT1 ~ PT3 from spot corrosion, test in U1 ~ U3 for 101 ~ 200 groups from uniform corrosion, and 201 ~ 300 groups from Inactivation experiment P1; Conveniently follow-up neural network classification, herein with 1 type represent spot corrosion, 2 types represent uniform corrosion, 3 types represent passivation; From known theory, passivation has maximum noise resistance Rn, and generally its order of magnitude is at 105 more than Ω, and the noise resistance of uniform corrosion and spot corrosion is then much smaller than this order of magnitude; The noise resistance value that Fig. 3 b shows 201 ~ 300 groups of samples is well about 105 Ω, is far longer than other samples, and other sample noise resistance are less than 104 Ω substantially; Theoretical according to shot noise, spot corrosion has larger feature charge value, and the feature charge value that Fig. 3 c demonstrates spot corrosion is far longer than other samples; After Fig. 3 e to Fig. 3 k represents calculating small echo normalized energy, the energy of every layer accounts for the ratio of gross energy; From wavelet decomposition theory, if wavelet coefficient is larger, then the similarity of it and morther wavelet is higher; Therefore transient state summit causes it to have larger wavelet coefficient at large scale layer, and high-frequency oscillation signal can have larger wavelet coefficient at small scale layer; Because spot corrosion signal exists a large amount of transient state peak, therefore it has larger energy distribution at large scale layer and low frequency layer; And uniform corrosion and unsharp signal are high-frequency oscillation signals, therefore it has larger energy distribution at small scale high frequency layer; Fig. 3-e ~ Fig. 3 k shows spot corrosion has larger energy value at d7 layer, and uniform corrosion has larger energy value at d1 ~ d3 layer, and passivation then has larger energy value at d4 ~ d6 layer; To sum up, the data result of Fig. 3 is consistent with erosion theory height;
Step 4) training of BP neural network and process:
By calculating the electrochemical signals eigenwert of each group, obtain the value of 300 groups of samples; 10 parameters often organizing sample are formed a stack features vector by the order of the energy Ratios of noise resistance, feature electric charge, characteristic frequency, small echo normalized energy d1 ~ d7 layer, and 300 groups of sample values can obtain 300 groups of different proper vectors; Every stack features vector, as the input set of BP neural network, often organizes the output set of corrosion type corresponding to sample as BP neural network; 60% input of getting input total value, as training set, gets the input of 20% as checking collection and test set; Use training set training data, do the selection of neural metwork training optimal value with checking collection, by Generalization Capability and the nicety of grading of test set test neural network; Concrete neural metwork training performance is as shown in Figure 4:
As shown in Figure 4, line B verifies that collection square error reduces along with the increase of iterations, and at the beginning of beginning, square error restrains rapidly and reduces, and achieves minimum value subsequently when iteration 26 times; If increase iterations again can increase training time and square error, be worthless; Therefore neural metwork training is achieving best checking performance after 26 iteration: 2.138 × 10
-6.
In step 4) in, the training of described BP neural network needs with process the optimization first carrying out BP neural network:
The target of BP neural network project training obtains a specific network, makes that the computing velocity of neural network is fast, training precision is high, generalization ability is strong; Computing velocity and neural network complete the time of sample training, and training precision refers to the training classification accuracy of training sample, and generalization ability refers to the accuracy rate of the neural network prediction classification trained.A kind of mutex relation between 3 targets of neural network, the fast network training precision of general computing velocity is low, the too high generalization ability of network energy of training precision is poor, and the network calculations speed that training precision is high is comparatively slow, therefore the relation how weighing three's target is a knowledge.
The optimizing process of BP neural network is then the process selecting most suitable three parameters: the sampling process of these three parameters of hidden layer nodes, learning rate and learning function is exactly the process of measurement three parameters.
Different learning function can affect training precision and generalization ability.The most frequently used function is have selected in the processing procedure of electrochemistry noise data
its Complete Classification training precision is high and generalization ability is strong.
Hidden layer nodes is very few has training speed but training precision also can be lower faster, and too much hidden layer nodes can cause over-fitting and slower training speed.Over-fitting and generalization ability poor, refer to the phenomenon that the precision of training sample is very high and the precision of forecast sample is low.The selection of hidden layer nodes can with reference to experimental formula: l≤n-1 or
wherein a is the value between 0 ~ 10.In electrochemical data process, the training through repeatedly neural network is compared, and finds performance the best when hidden layer nodes selects n-1.
Learning rate refers to the speed of convergence of weights, larger learning rate can strengthen the size of each right value update, but the concussion of weights may be caused simultaneously, iterations is increased greatly and even cannot form convergence, and less learning rate can make the value of each right value update very little, lengthen the time of training.In the process of electrochemical data process, we have employed the method becoming learning rate, make initial weight rapidly near optimal weight, adopt the generation that less learning rate prevents weight from shaking afterwards at the beginning of starting training with larger learning rate.
Claims (2)
1. based on an electrochemistry noise corrosion signal processing method for neural network, it is characterized in that: the described corrosion of the electrochemistry noise based on neural network signal processing method comprises the following step carried out in order:
Step 1) refine the typical electrochemical noise signal of different corrosion type:
Corrosion region is divided into spot corrosion, uniform corrosion and passivation three kinds of corrosion types by the time-domain diagram according to the electrochemical noise signal obtained from electrochemistry noise experiment; Described electrochemistry noise experiment makes working electrode with standard 304 stainless steel sample, and two working electrodes and a contrast electrode are put among experimental solutions, measure the electrochemical source of current noise between two working electrodes and the electrochemical voltage noise between working electrode and contrast electrode by zero resistance galvanometer simultaneously; The electrochemistry noise of its performance of different corrosion types has very large difference on time-domain diagram; When there is spot corrosion, current noise and voltage noise there will be obvious transient state peak; When there is passivation or uniform corrosion, current noise and voltage noise then show as the irregular vibration of high frequency;
Step 2) packet:
Be that a unit carries out data cutting by the above-mentioned experimental data collected by 1024 points, namely within every 512 seconds, deposit an independent small documents, filename temporally superposes, then obtain the data of a large amount of various corrosion;
Step 3) often organized each characteristic parameter of data:
To often kind of corrosion type random selecting up to a hundred groups of data in these all small documents, these data of batch processing, are often organized the value of each characteristic parameter of data;
Step 4) training of BP neural network and process:
By calculating the electrochemical signals eigenwert of each group, obtain the value of up to a hundred groups of samples; 10 parameters often organizing sample are formed a stack features vector by the order of the energy Ratios of noise resistance, feature electric charge, characteristic frequency, small echo normalized energy d1 ~ d7 layer, and up to a hundred groups of sample values can obtain up to a hundred groups of different proper vectors; Every stack features vector, as the input set of BP neural network, often organizes the output set of corrosion type corresponding to sample as BP neural network; 60% input of getting input total value, as training set, gets the input of 20% as checking collection and test set; Use training set training data, do the selection of neural metwork training optimal value with checking collection, by Generalization Capability and the nicety of grading of test set test neural network.
2. the corrosion of the electrochemistry noise based on neural network signal processing method according to claim 1, is characterized in that: in step 4) in, need the optimization first carrying out BP neural network before the training of described BP neural network and process:
The optimizing process of BP neural network is the process selecting most suitable three parameters: hidden layer nodes, learning rate and learning function;
Select the most frequently used function
its Complete Classification training precision is high and generalization ability is strong;
The selection of hidden layer nodes with reference to experimental formula: l≤n-1 or
wherein a is the value between 0 ~ 10.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410847805.XA CN104568730A (en) | 2014-12-31 | 2014-12-31 | Electrochemical noise corrosion signal processing method based on neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410847805.XA CN104568730A (en) | 2014-12-31 | 2014-12-31 | Electrochemical noise corrosion signal processing method based on neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104568730A true CN104568730A (en) | 2015-04-29 |
Family
ID=53085308
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410847805.XA Pending CN104568730A (en) | 2014-12-31 | 2014-12-31 | Electrochemical noise corrosion signal processing method based on neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104568730A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101303329A (en) * | 2008-06-13 | 2008-11-12 | 东南大学 | Integrated strength testing method based on neural network technology |
| CN101329169A (en) * | 2008-07-28 | 2008-12-24 | 中国航空工业第一集团公司北京航空制造工程研究所 | Neural network modeling approach of electron-beam welding consolidation zone shape factor |
| CN101865817A (en) * | 2010-06-08 | 2010-10-20 | 天津大学 | Sensor and detection method for detecting corrosion of buried metal |
| CN103439631A (en) * | 2013-08-12 | 2013-12-11 | 国家电网公司 | Method and system for detecting corrosion state of grounding grid |
-
2014
- 2014-12-31 CN CN201410847805.XA patent/CN104568730A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101303329A (en) * | 2008-06-13 | 2008-11-12 | 东南大学 | Integrated strength testing method based on neural network technology |
| CN101329169A (en) * | 2008-07-28 | 2008-12-24 | 中国航空工业第一集团公司北京航空制造工程研究所 | Neural network modeling approach of electron-beam welding consolidation zone shape factor |
| CN101865817A (en) * | 2010-06-08 | 2010-10-20 | 天津大学 | Sensor and detection method for detecting corrosion of buried metal |
| CN103439631A (en) * | 2013-08-12 | 2013-12-11 | 国家电网公司 | Method and system for detecting corrosion state of grounding grid |
Non-Patent Citations (1)
| Title |
|---|
| 孔伟康: "基于电化学噪声的腐蚀信号的数据处理技术的研究", 《万方学位论文全文数据库》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106502092B (en) | A kind of thermal process model parameter identification method using improvement Hybrid Particle Swarm | |
| Pristyanto et al. | Data level approach for imbalanced class handling on educational data mining multiclass classification | |
| Chen et al. | Solving many-objective optimization problems via multistage evolutionary search | |
| CN109033730A (en) | A kind of tool wear prediction technique based on improved particle swarm optimization algorithm | |
| CN109462257B (en) | Sensitivity identification method considering voltage stability of multivariate random variable power grid | |
| CN107944559A (en) | A kind of entity relationship automatic identifying method and system | |
| CN104933156A (en) | Collaborative filtering method based on shared neighbor clustering | |
| Zhou et al. | Ranking vectors by means of the dominance degree matrix | |
| CN111046327B (en) | Prony analysis method suitable for low-frequency oscillation and subsynchronous oscillation identification | |
| Khan et al. | Novel approach for a van der pol oscillator in the continuous time domain | |
| CN110987436B (en) | Bearing fault diagnosis method based on excitation mechanism | |
| CN114492675B (en) | Intelligent fault cause diagnosis method for capacitor voltage transformer | |
| CN114036610A (en) | Penetration depth prediction method based on data enhancement | |
| CN112699596A (en) | Wide-area electromagnetic method induced polarization information nonlinear extraction method based on learning | |
| CN116796403A (en) | Building energy saving method based on comprehensive energy consumption prediction of commercial building | |
| Bahmanikashkooli et al. | Application of particle swarm optimization algorithm for computing critical depth of horseshoe cross section tunnel | |
| CN111274701A (en) | Harmonic source affine modeling method adopting interval monitoring data dimension reduction regression | |
| WO2022032874A1 (en) | Adversarial neural network-based hydrological parameter calibration method for data region | |
| CN104568730A (en) | Electrochemical noise corrosion signal processing method based on neural network | |
| Zhao et al. | A Back Propagation Neural Network Model based on kalman filter for water quality prediction | |
| CN111260029A (en) | Credibility analysis method for air quality data | |
| Zheng et al. | Uncertainty Quantification Analysis on Silicon Electrodeposition Process Via Numerical Simulation Methods | |
| Zhou et al. | General regression neural network forecasting model based on PSO algorithm in water demand | |
| Reddy et al. | Practical model for predicting beta transus temperature of titanium alloys | |
| CN104502267A (en) | Electrochemical noise corrosion type distinguishing method based on support vector machine |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150429 |