CN109444360A - Juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction - Google Patents
Juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction Download PDFInfo
- Publication number
- CN109444360A CN109444360A CN201811270581.5A CN201811270581A CN109444360A CN 109444360 A CN109444360 A CN 109444360A CN 201811270581 A CN201811270581 A CN 201811270581A CN 109444360 A CN109444360 A CN 109444360A
- Authority
- CN
- China
- Prior art keywords
- neural network
- template
- cell
- sensor
- cell neural
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/02—Food
- G01N33/14—Beverages
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Food Science & Technology (AREA)
- Chemical & Material Sciences (AREA)
- Medicinal Chemistry (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Investigating Or Analyzing Materials By The Use Of Electric Means (AREA)
Abstract
本发明公开了一种基于细胞神经网络和电子鼻特征提取的果汁贮藏期检测算法，获取果汁气体样本对果汁气体样本进行气味检测，得到果汁气体原始响应数据集；筛选传感器，得到响应数据集和响应矩阵；选择模板，根据响应矩阵，建立带未知参数的果汁气体细胞神经网络；将标记响应数据集输入到带未知参数的果汁气体细胞神经网络中进行特征点标记，得到传感器特征点；寻求未知参数最优值，并对所有传感器特征点进行特征提取，得出对应的细胞神经网络模板，以及对应的细胞神经网络；根据传感器特征点对果汁气体细胞神经网络进行验证。有益效果：适用范围广，对果汁产品储存器进行实时监控，便于监管部分检查，对果汁品质进行鉴别。保证果汁行业质量。
The invention discloses a juice storage period detection algorithm based on cellular neural network and electronic nose feature extraction. The juice gas sample is obtained and the juice gas sample is odor detected to obtain a juice gas original response data set; sensors are screened to obtain a response data set and a Response matrix; select the template, according to the response matrix, build a juice gas cell neural network with unknown parameters; input the marked response data set into the juice gas cell neural network with unknown parameters to mark feature points to obtain sensor feature points; seek unknown parameters The optimal value of the parameters is obtained, and feature extraction is performed on all sensor feature points to obtain the corresponding cellular neural network template and the corresponding cellular neural network; the juice gas cellular neural network is verified according to the sensor feature points. Beneficial effects: Wide range of application, real-time monitoring of juice product storage, convenient for supervision part inspection, and identification of juice quality. Guarantee the quality of juice industry.
Description
Technical field
The present invention relates to juice storage phase detection technique fields, specifically a kind of to be based on cell neural network and electronics
The juice storage phase detection algorithm of nose feature extraction.
Background technique
Hyundai electronics nasus system usually has the sensor array comprising dozens of gas sensor, in a gas sample
Quite huge sample data can be obtained in acquisition, and because of the cross-sensitivity of electronic nose sensor array, difference sensing
Device can all respond identical gas, and the result for directly obtaining their input pattern recognizers is very undesirable.
Feature extraction is the first step of sensor signal processing, it plays an important role in subsequent pattern-recognition: from
Prior feature is extracted in electronic nose original response, facilitates Lifting scheme accuracy of identification.Most basic feature extracting method
It is direct processing original response curve, is such as maximized, minimum value, slope, the corresponding time coordinate of max min, integrates
Deng.These values composition primitive character matrix is used for pattern-recognition.The wherein end-state of maximum value representative sensor response,
It is to be most commonly used for distinguishing the steady state characteristic that gas is answered.On the basis of these methods, better effect in order to obtain, one
A little classical algorithms be used to handle primitive character matrix again.For example, principal component analysis is that one kind is intended to find out several synthesis
Index reflects the information of primal variable to represent numerous features, these overall targets originally as much as possible, and each other mutually
Incoherent method.However, steady-state response is not good enough for effect in complicated pattern-recognition, because it often has ignored many
Important information.Therefore, feature extraction is also required to behavioral characteristics.
When to the fruit storage phase, when the content detection of fruit syrup, due to the general different area of fruit, and it is exported to the world
Various regions, since each department weather, environment are different, fruit trade company needs to carry out testing and evaluation to the quality of fruit in real time, thus
Specified storage plan and marketing plan, however in the prior art, there are no the methods for carrying out intellectualized detection to fruit syrup, also
It rests on the mode artificially inspected by random samples to be detected, sampling observation process is complicated, and needs long-term existence individual difference, obtained conclusion essence
Exactness is low, is unable to meet demand.
Summary of the invention
In view of the above-mentioned problems, being store the present invention provides a kind of based on cell neural network and the fruit juice of electronic nose feature extraction
Hiding phase detection algorithm is measured in real time the fruit juice during fruit storage, qualitative change or lesion occurs, and improves fruit and saves
Reliability.
In order to achieve the above objectives, the specific technical solution that the present invention uses is as follows:
A kind of juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction, key are:
S1: fruit juice gas sample is obtained；
S2: building fruit juice gas electronic nose monitors system, and carries out odor detection to fruit juice gas sample, obtains fruit juice gas
Body original response data set；
S3: screening sensor；Sensor is screened using X binary parameters, obtains labeled sensor group, the sensing
The collected data of device group are label response data sets, and are arranged in response matrix M × N；
S4: cell neural network template of the selection with unknown parameter establishes band unknown parameter according to response matrix is arranged in
Fruit juice gas cell neural network；The label response data sets of step S3 are input to the fruit juice gas cell with unknown parameter
Characteristic point label is carried out in neural network, obtains sensor characteristics point；
S5: using enhanced krill colony optimization algorithm, seeks optimal value to all unknown parameters, and to all the sensors spy
Sign point carries out feature extraction, obtains corresponding cell neural network template and corresponding cell neural network；
S6: the classifying quality obtained according to sensor characteristics point tests the ability in feature extraction of cell neural network
Card.
Further, it obtains fruit juice gas sample in step sl to be arranged in vaporizer, which includes fruit juice
Container, which is provided with venthole and one end of appendix connects, and the other end of the appendix protrudes into the electronics
Nose monitors in the monitoring chamber of system.
Wherein, when making fruit juice, the fruit of the same similar maturity of arbitrary extracting X kg；It is squeezed the juice, is passed through using cold press technique
It is packed into vaporizer after filtering, sterilizing；Every N minutes stirring is primary.Fruit gas is sent into electronic nose monitoring system from appendix
It monitors in chamber.
Further, the sensor in fruit juice gas electronic nose monitoring system in step s 2 needs to pre-process:
S21: being exposed to clean air Q minutes for all the sensors, to obtain detection baseline；
S22: fruit juice gas sample is introduced in the monitoring chamber of the electronic nose monitoring system, the duration is P minutes；
Return step S21.Q, P is set according to actual needs.
Before each sampling test, self-test is carried out to fruit juice gas electronic nose monitoring system, to the power of sensor array
Then the preheated one-section time carries out normal odor detection.In the detection process, it is sensed using smell acquisition software real-time display
Response curve of the device array to orange juice smell.
Wherein, the environmental factor in setting fruit juice gas electronic nose monitoring system is also needed.Wherein at least include environment temperature,
Humidity.
Further, when screening sensor, when setting the binary number of corresponding sequence digit as 1, then labeled as choosing
In the sensor；Labeled as 0, then labeled as unselected.
Further, the fruit juice gas cell neural network in step S4 with unknown parameter are as follows:
I=1,2 ..., M, j=1,2 ..., N；
xij(t) be cell neural network any cell C (i, j) state value, primary condition xij(0)=0；
uijIt is cell neural network any cell C (i, j) input, wherein static input | uij|≤1；
yij(t) be cell neural network any cell C (i, j) output；
A(i,j；k,l),B(i,j；K, l), I be cell neural network cell C (k, l) C (i, j) template；
A is a feedback template, and B is a control cloned template, and I is threshold value, is threshold value unknown parameter；
The state of each unit and output are related to the unit for being connected to it, for each cell C (i, j), below
Set Nr(i, j) can be defined as name R neighborhood；
Nr(i,j)
=C (k.l) :=max (| (k-i) |, | l-j |)≤r, 1≤k≤M, 1≤l≤N,
Wherein, radius of neighbourhood r is positive integer, is the index for indicating cell position, i.e. GU Generic Unit to (i, j) and (k, l)
And its row and column of adjacent cells within a grid.
Therefore, for cell arrangement in the R neighborhood of C (i, j), size is (2r+1) × (2r+1) grid.
Further, in step S4, cell neural network type includes that single mode plate cell neural network and multi-template are thin
Born of the same parents' neural network；
When for single mode plate cell neural network:
When multi-template cell neural network:
Wherein i is the sequence number of sensor, and j is j-th of sampled point of i-th of sensor；I=1,2 ... M；M≤X；J=
1,2,…1024；When X is screening sensor, the digit of binary number；M is template number, equal with the sensor number of screening；
'+' and '-' are for marking time sequencing；
x1、x2、x3, b be control the unknown control parameter of cloned template；A is the unknown control parameter for feeding back template.
Further, it in step S4, feeds back unknown in unknown control parameter a, the control cloned template B in template A
Control parameter x1、x2、x3, b, Existence restraint condition between threshold value I:
Further, in step s 5, unknown parameter is selected using enhanced krill colony optimization algorithm,
In, unknown parameter includes feeding back the unknown control parameter a in template A, the unknown control parameter x in control cloned template B1、x2、
x3, b, threshold value I；
The specific steps of step S5 are as follows:
Step (a): sensor representative was carried out using the X bit generated by enhanced krill colony optimization algorithm
Filter；
Step (b): the response that sensor obtains is aligned in response matrix M × N of cell neural network input；
Step (c): if selection multi-template cell neural network, response matrix M × N input cell neural network, make to ring
Answer the corresponding template of any sensor of matrix M × N；
If selecting single mode plate, all the sensors are used for using the same template；The parameter of the same template is by enhanced
Krill colony optimization algorithm optimizes；
Step (d): the sensor characteristics point of cell neural network is handled, to obtain feature extraction result；
Step (e): feature extraction result is input in classifier to obtain discrimination, and excellent using enhanced krill group
Change the parameter that algorithm carrys out Optimum Classification device；
Step (f): it repeats the above steps to obtain overall optimal parameter；It obtains and is most suitable for what each sensor characteristics were extracted
Cell neural network template.
Further, enhanced krill colony optimization algorithm step are as follows:
Step 1: initialization initializes the number of iterations, enables I=1；And NP krill population P is initialized, set speed of looking for food
Vf, maximally diffuse speed Dmax, maximum induced velocity Nmax, maximum number of iterations MI；
Step 2: fitness calculates；Its grade of fit is calculated according to the initial position of each krill krill population；
Step 3: as I < MI, all krill populations are ranked up according to grade of fit；And to krill population execute with
Lower motion calculation:
It is looked for food using the krill population movement of other individual inductions, realizes movement physical diffusion；
According to formulaCalculate decision weights factor dxi/dt；
Ni, Fi and Di indicate movement of looking for food, and the physical diffusion by other krill populations and krill population i is influenced；The
One movement Fi includes two parts: current foodstuff position and the information in relation to previous position；Realize crossing operation symbol；
Update the position of the krill population in search space；And it is calculated according to the new position of krill population and new is suitble to
Degree；I=I+1 is enabled, step 1 is returned.
Beneficial effects of the present invention: it experiments have shown that the behavioral characteristics of electronic nose sensor response have more distinction, is input to
Effect in classifier is more preferable than traditional characteristic extracting method.Cell neural network devises the template applied to electric nasus system
With data preprocessing method, solve the problems, such as that the electronic nose sensor response of time series is not available cell neural network.
This method can be applied to various fruit and fruit juice, and in addition to the detection of storage period, it can be also used for detecting different processing classes
Type, fruit juice concn and other information, this is conducive to following fruit classification fruit juice production industry quality.It may be supervisor
The quality inspection of structure provides reference.
Detailed description of the invention
The present invention is based on the feature extraction flow charts of cell neural network (CNN) by Fig. 1；
Fig. 2 experimental system schematic diagram；
The response of Fig. 3 inventive sensor array；
Fig. 4 two dimension CNN structure chart；
Fig. 5 CNN system structure；
Fig. 6 CNN output function；
The processing schematic of Fig. 7 proposal template；
The accuracy of identification (%) of different characteristic point after Fig. 8 CNN processing；
The recognition accuracy (%) of Fig. 9 different characteristic extractive technique；
The recognition accuracy (%) of Figure 10 single mode plate and multi-template CNN.
Specific embodiment
Specific embodiment and working principle of the present invention will be described in further detail with reference to the accompanying drawing.
A kind of juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction, can be in conjunction with Fig. 1
Find out, committed step are as follows:
S1: fruit juice gas sample is obtained；In the present embodiment, it by taking orange blossom as an example, is analyzed.
It obtains fruit juice gas sample in step sl to be arranged in vaporizer, which includes juice container, the fruit juice
Container is provided with venthole and one end of appendix connects, and the other end of the appendix protrudes into the electronic nose monitoring system
It monitors in chamber.
Wherein, when making fruit juice, the fruit of the same similar maturity of arbitrary extracting 5kg；It is squeezed the juice, is passed through using cold press technique
It is packed into vaporizer after filter, sterilizing；Stirring in every 15 minutes is primary.Fruit gas is sent into the prison of electronic nose monitoring system from appendix
It surveys in chamber.
As it can be seen from table 1 as shown in table 1 to the analysis of orange blossom fragrance component.
1 orange blossom analysis of aroma components table of table
S2: building fruit juice gas electronic nose monitors system, and carries out odor detection to fruit juice gas sample, obtains fruit juice gas
Body original response data set；
The sensor in fruit juice gas electronic nose monitoring system in step s 2 needs to pre-process:
S21: being exposed to clean air 5 minutes for all the sensors, to obtain detection baseline；
S22: fruit juice gas sample is introduced in the monitoring chamber of the electronic nose monitoring system, the duration is 7 minutes；
Return step S21.
Wherein, in fruit juice gas electronic nose monitoring system, see Table 2 for details for the sensitive features distribution of each gas sensor.
The sensitive prime characteristic of 2 gas sensor of table
Note: the response of these sensors is nonspecific.Other than the sensitive prime gas in table 2, they may be also
To other gas sensitizations.
Wherein, the temperature and humidity of gas electronic nose monitoring system detection chamber is set as 25 DEG C and 40% first.
Further, when screening sensor, when setting the binary number of corresponding sequence digit as 1, then labeled as choosing
In the sensor；Labeled as 0, then labeled as unselected.
S3: screening sensor；Sensor is screened using 15 binary parameters, obtains labeled sensor group, the biography
The collected data of sensor group are label response data sets, and are arranged in response matrix M × N；It is detailed in Fig. 3
S4: cell neural network template of the selection with unknown parameter establishes band unknown parameter according to response matrix is arranged in
Fruit juice gas cell neural network；The label response data sets of step S3 are input to the fruit juice gas cell with unknown parameter
Characteristic point label is carried out in neural network, obtains sensor characteristics point；
Fig. 4 shows a simple two dimension CNN structure, and square indicates unity element, and the line between unit indicates single
Interaction between member.As far as we know, each unit is connected solely to the nearest-neighbors in network, and adjacent cell is straight
Connect interaction.Due to the duration dynamic transmission of CNN, there is indirect influence between the cell not being directly connected to.It is any to choose
One representational cell, which is connected thereto, under, left and right, upper left, lower-left, upper right and lower right list
First lattice.The cell at 4 angles will be connected to 3 adjacent cells, and remaining peripheral cells will be connected to 5 cells.Each cell
It is a nonlinear dynamic system, state is related with input, output and dynamics rule.
Further, the fruit juice gas cell neural network in step S4 with unknown parameter are as follows:
I=1,2 ..., M, j=1,2 ..., N；
xij(t) be cell neural network any cell C (i, j) state value, primary condition xij(0)=0；
uijIt is cell neural network any cell C (i, j) input, wherein static input | uij|≤1；
yij(t) be cell neural network any cell C (i, j) output；
A(i,j；k,l),B(i,j；K, l), I be cell neural network cell C (k, l) C (i, j) template；
A is a feedback template, and B is a control cloned template, and I is threshold value, is threshold value unknown parameter；
The state of each unit and output are related to the unit for being connected to it, for each cell C (i, j), below
Set Nr(i, j) can be defined as name R neighborhood；
Nr(i,j)
=C (k.l) :=max (| (k-i) |, | l-j |)≤r, 1≤k≤M, 1≤l≤N,
Wherein, radius of neighbourhood r is positive integer, is the index for indicating cell position, i.e. GU Generic Unit to (i, j) and (k, l)
And its row and column of adjacent cells within a grid.
Therefore, for cell arrangement in the R neighborhood of C (i, j), size is (2r+1) × (2r+1) grid.Cytocidal action
The working principle of network is as shown in figure 5, output function is as shown in Figure 6.
In step S4, cell neural network type includes single mode plate cell neural network and multi-template cell neural network；
When for single mode plate (single-template) cell neural network:
When multi-template (multi-template) cell neural network:
Wherein i is the sequence number of sensor, as shown in fig. 7, j is j-th of sampled point of i-th of sensor；
I=1,2 ... M；M≤X；J=1,2 ... 1024；When X is screening sensor, the digit of binary number；M is template
Number is equal with the sensor number of screening；
'+' and '-' are for marking time sequencing；
x1、x2、x3, b be control the unknown control parameter of cloned template；A is the unknown control parameter for feeding back template.
In conjunction with table 3, in step S4, the unknown control parameter a in template A, the unknown control in control cloned template B are fed back
Parameter x1、x2、x3, b, Existence restraint condition between threshold value I:
Further, in step s 5, unknown parameter is selected using enhanced krill colony optimization algorithm,
In, unknown parameter includes feeding back the unknown control parameter a in template A, the unknown control parameter x in control cloned template B1、x2、
x3, b, threshold value I；
The details for the parameter that table 3 needs to optimize
S5: using enhanced krill colony optimization algorithm, seeks optimal value to all unknown parameters, and to all the sensors spy
Sign point carries out feature extraction, obtains corresponding cell neural network template and corresponding cell neural network；The tool of step S5
Body step are as follows:
Step (a): sensor representative was carried out using the X bit generated by enhanced krill colony optimization algorithm
Filter；
Step (b): the response that sensor obtains is aligned in response matrix M × N of cell neural network input；
Step (c): if selection multi-template cell neural network, response matrix M × N input cell neural network, make to ring
Answer the corresponding template of any sensor of matrix M × N；
If selecting single mode plate, all the sensors are used for using the same template；The parameter of the same template is by enhanced
Krill colony optimization algorithm optimizes；
Step (d): the sensor characteristics point of cell neural network is handled, to obtain feature extraction result；
In conjunction with Fig. 7 as can be seen that have multiple characteristic points labeled when marker characteristic point, such as density bullet in Fig. 7
Point is labeled characteristic point, but ideally, is that these labeled characteristic points are represented with a value, in conjunction with Fig. 8,
Using the maximum value for calculating all characteristic points, minimum value, average value and difference, maximum value, minimum value, average value and difference are exactly special
Sign extracts result.
Step (e): feature extraction result is input in classifier to obtain discrimination, and excellent using enhanced krill group
Change the parameter that algorithm carrys out Optimum Classification device；
The orange blossom used in the present embodiment has 4 kinds of storage periods, so each sample corresponds to 4 kinds of different labels 1
23 4. these sample datas, either training data and test data input after classifier, and classifier judges out them
Label, the label that classifier is judged and true tag comparison, can obtain discrimination.And it is excellent using enhanced krill group
Change the parameter that algorithm carrys out Optimum Classification device.
Step (f): it repeats the above steps to obtain overall optimal parameter；It obtains and is most suitable for what each sensor characteristics were extracted
Cell neural network template.
Further, enhanced krill colony optimization algorithm step are as follows:
Step 1: initialization initializes the number of iterations, enables I=1；And NP krill population P is initialized, set speed of looking for food
Vf, maximally diffuse speed Dmax, maximum induced velocity Nmax, maximum number of iterations MI；
Step 2: fitness calculates；Its grade of fit is calculated according to the initial position of each krill krill population；
Step 3: as I < MI, all krill populations are ranked up according to grade of fit；And to krill population execute with
Lower motion calculation:
It is looked for food using the krill population movement of other individual inductions, realizes movement physical diffusion；
According to formulaCalculate decision weights factor dxi/dt；
Ni, FiAnd DiExpression is looked for food movement, and the physical diffusion by other krill populations and krill population i is influenced；First
A movement FiIncluding two parts: current foodstuff position and in relation to the information of previous position；Realize crossing operation symbol；
Update the position of the krill population in search space；And it is calculated according to the new position of krill population and new is suitble to
Degree；I=I+1 is enabled, step 1 is returned.
S6: the classifying quality obtained according to sensor characteristics point tests the ability in feature extraction of cell neural network
Card.
The feature extraction result and support vector machines (SVM) of CNN, radial basis function (RBF), core linear discriminant analysis
(KLDA) three kinds of classical taxonomy devices combine.Enhanced krill colony optimization algorithm (EKH) is used for parameter optimization.Each program repeats
10 times, wherein optimum is the final result of each classifier, and the results are shown in Table 4, describes test data (Acc_
Test) and training data (Acc_train), it is evident that when being marked with CNN, characteristic point has good dynamic characteristic with area
Divide the Storage period of orange blossom.Fig. 8 shows that test data concentrates the recognition accuracy of different value.Average value is than other processing methods
More comprehensively, the dynamic characteristic and optimum performance that can be responded with representative sensor.
Therefore, compared with other feature extracting methods, average value is derived from feature as sensor response characteristic.Such as preceding institute
It states, they are in response to maximum value (Max), integrate (Integral), curve matching (Curve Fitting), principal component analysis
(PCA) and core principle component analysis (KPCA).Then we use support vector machines (SVM), radial basis function neural network
(RBFNN), the effect of these feature extractions is compared in core linear discriminant analysis (KLDA) as classifier.The results are shown in Table 5.
As can be seen that CNN feature extraction effect is substantially better than other algorithms, as shown in Figure 9.
The nicety of grading of different characteristic point shows table (%) after the processing of 4 cell neural network of table
The classification accuracy (%) that 5 different characteristic of table is extracted
It is shown above the result is that being obtained from multi-template CNN.In addition, the performance test of single mode plate CNN is as follows.Single mode plate
It is as shown in table 6 with the recognition effect of multi-template CNN.
It may be seen that single template version of CNN can be a bit weaker, but compared with other schemes shown in Fig. 10, it is still
With good ability in feature extraction.
The result shows that CNN can be applied to electronic nose feature extraction field, it is proposed by the present invention that there is good identification essence
Degree.Again it is apparent that CNN outstanding image-capable can be applied to electronic nose feature extraction field, the spy based on CNN
Sign extraction algorithm helps to improve the accuracy of electronic nose identification.
The classification accuracy (%) of 6 single mode plate of table and multi-template CNN
It should be pointed out that the above description is not a limitation of the present invention, the present invention is also not limited to the example above,
Variation, modification, addition or the replacement that those skilled in the art are made within the essential scope of the present invention, are also answered
It belongs to the scope of protection of the present invention.
Claims (9)
1. a kind of juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction, it is characterised in that:
S1: fruit juice gas sample is obtained；
S2: building fruit juice gas electronic nose monitors system, and carries out odor detection to fruit juice gas sample, and it is former to obtain fruit juice gas
Beginning response data sets；
S3: screening sensor；Sensor is screened using X binary parameters, obtains labeled sensor group, the sensor group
Collected data are label response data sets, and are arranged in response matrix M × N；
S4: cell neural network template of the selection with unknown parameter establishes the fruit with unknown parameter according to response matrix is arranged in
Juice gas cell neural network；The label response data sets of step S3 are input to the fruit juice gas cellular neural with unknown parameter
Characteristic point label is carried out in network, obtains sensor characteristics point；
S5: enhanced krill colony optimization algorithm is used, optimal value is sought to all unknown parameters, and to all the sensors characteristic point
Feature extraction is carried out, obtains corresponding cell neural network template and corresponding cell neural network；
S6: the classifying quality obtained according to sensor characteristics point verifies the ability in feature extraction of cell neural network.
2. according to claim 1 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that: fruit juice gas sample is obtained in step sl to be arranged in vaporizer, which includes juice container,
The juice container is provided with venthole and one end of appendix connects, and the other end of the appendix protrudes into the electronic nose monitoring
In the monitoring chamber of system.
3. according to claim 1 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that: the sensor in fruit juice gas electronic nose monitoring system in step s 2 needs to pre-process:
S21: being exposed to clean air Q minutes for all the sensors, to obtain detection baseline；
S22: fruit juice gas sample is introduced in the monitoring chamber of the electronic nose monitoring system, the duration is P minutes；It returns
Step S21.
4. according to claim 1 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that: when screening sensor, when setting the binary number of corresponding sequence digit as 1, then labeled as choosing the biography
Sensor；Labeled as 0, then labeled as unselected.
5. according to claim 1 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that: the fruit juice gas cell neural network in step S4 with unknown parameter are as follows:
I=1,2 ..., M, j=1,2 ..., N；
xij(t) be cell neural network any cell C (i, j) state value, primary condition xij(0)=0；
uijIt is cell neural network any cell C (i, j) input, wherein static input | uij|≤1；
yij(t) be cell neural network any cell C (i, j) output；
A(i,j；k,l),B(i,j；K, l), I be cell neural network cell C (k, l) C (i, j) template；
A is a feedback template, and B is a control cloned template, and I is threshold value, is threshold value unknown parameter.
6. according to claim 5 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that: in step S4, cell neural network type includes single mode plate cell neural network and multi-template cellular neural
Network；
When for single mode plate cell neural network:
I=Ii,j
When multi-template cell neural network:
Wherein i is the sequence number of sensor, and j is j-th of sampled point of i-th of sensor；I=1,2 ... M；M≤X；J=1,
2,…1024；When X is screening sensor, the digit of binary number；M is template number, equal with the sensor number of screening；
'+' and '-' are for marking time sequencing；
x1、x2、x3, b be control the unknown control parameter of cloned template；A is the unknown control parameter for feeding back template.
7. according to claim 5 or 6 detected based on the juice storage phase of cell neural network and electronic nose feature extraction
Algorithm, it is characterised in that in step S4, feed back the unknown control parameter a in template A, the unknown control in control cloned template B
Parameter x1、x2、x3, b, Existence restraint condition between threshold value I:
8. according to claim 7 detected based on the juice storage phase of cell neural network and electronic nose feature extraction is calculated
Method, it is characterised in that in step s 5, unknown parameter is selected using enhanced krill colony optimization algorithm, wherein unknown
Parameter includes feeding back the unknown control parameter a in template A, the unknown control parameter x in control cloned template B1、x2、x3, b, threshold
Value I；
The specific steps of step S5 are as follows:
Step (a): sensor representative is filtered using the X bit generated by enhanced krill colony optimization algorithm；
Step (b): the response that sensor obtains is aligned in response matrix M × N of cell neural network input；
Step (c): if selection multi-template cell neural network, response matrix M × N input cell neural network, make to respond square
The corresponding template of any sensor of battle array M × N；
If selecting single mode plate, all the sensors are used for using the same template；The parameter of the same template is by enhanced krill
Colony optimization algorithm optimizes；
Step (d): the sensor characteristics point of cell neural network is handled, to obtain feature extraction result；
Step (e): feature extraction result is input in classifier to obtain discrimination, and calculated using enhanced krill group optimization
Method carrys out the parameter of Optimum Classification device；
Step (f): it repeats the above steps to obtain overall optimal parameter；It obtains and is most suitable for the cell that each sensor characteristics are extracted
Neural network template.
9. according to claim 1 or 8 based on the juice storage phase of cell neural network and electronic nose feature extraction detect
Algorithm, it is characterised in that enhanced krill colony optimization algorithm step are as follows:
Step 1: initialization initializes the number of iterations, enables I=1；And NP krill population P is initialized, set the speed V that looks for foodf, maximum
Diffusion velocity Dmax, maximum induced velocity Nmax, maximum number of iterations MI；
Step 2: fitness calculates；Its grade of fit is calculated according to the initial position of each krill krill population；
Step 3: as I < MI, all krill populations are ranked up according to grade of fit；And following fortune is executed to krill population
It is dynamic to calculate:
It is looked for food using the krill population movement of other individual inductions, realizes movement physical diffusion；
According to formulaCalculate decision weights factor dxi/dt；
Ni, FiAnd DiExpression is looked for food movement, and the physical diffusion by other krill populations and krill population i is influenced；First dynamic
Make FiIncluding two parts: current foodstuff position and in relation to the information of previous position；Realize crossing operation symbol；
Update the position of the krill population in search space；And its new grade of fit is calculated according to the new position of krill population；Enable I
=I+1 returns to step 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811270581.5A CN109444360B (en) | 2018-10-29 | 2018-10-29 | Juice storage period detection algorithm based on cellular neural network and electronic nose feature extraction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811270581.5A CN109444360B (en) | 2018-10-29 | 2018-10-29 | Juice storage period detection algorithm based on cellular neural network and electronic nose feature extraction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109444360A true CN109444360A (en) | 2019-03-08 |
| CN109444360B CN109444360B (en) | 2021-05-04 |
Family
ID=65549101
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811270581.5A Active CN109444360B (en) | 2018-10-29 | 2018-10-29 | Juice storage period detection algorithm based on cellular neural network and electronic nose feature extraction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109444360B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112433028A (en) * | 2020-11-09 | 2021-03-02 | 西南大学 | Electronic nose gas classification method based on memristor cell neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105891430A (en) * | 2016-04-11 | 2016-08-24 | 上海大学 | Food quality monitoring device based on neural network |
| CN108304970A (en) * | 2018-02-05 | 2018-07-20 | 西北农林科技大学 | The method for quick predicting and system of Apple, air conditioned storage monitoring system |
-
2018
- 2018-10-29 CN CN201811270581.5A patent/CN109444360B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105891430A (en) * | 2016-04-11 | 2016-08-24 | 上海大学 | Food quality monitoring device based on neural network |
| CN108304970A (en) * | 2018-02-05 | 2018-07-20 | 西北农林科技大学 | The method for quick predicting and system of Apple, air conditioned storage monitoring system |
Non-Patent Citations (5)
| Title |
|---|
| LEON 0. CHUA等: "Cellular Neural Networks: Theory", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS》 * |
| LI WANG等: "A Novel Optimization Technique to Improve Gas Recognition by Electronic Noses Based on the Enhanced Krill Herd Algorithm", 《SENSORS》 * |
| PIOTR A. KOWALSKI等: "Training Neural Networks with Krill Herd Algorithm", 《NEURAL PROCESS LETT》 * |
| TUBA AYHAN等: "AN APPLICATION OF SMALL-WORLD CELLULAR NEURAL NETWORKS ON ODOR CLASSIFICATION", 《INTERNATIONAL JOURNAL OF BIFURCATION AND CHAOS》 * |
| 张虹艳等: "基于BP神经网络的电子鼻羊奶贮藏时间的预测", 《食品工业科技》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112433028A (en) * | 2020-11-09 | 2021-03-02 | 西南大学 | Electronic nose gas classification method based on memristor cell neural network |
| CN112433028B (en) * | 2020-11-09 | 2021-08-17 | 西南大学 | Electronic nose gas classification method based on memristive cellular neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109444360B (en) | 2021-05-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sadeghi-Tehran et al. | Automated method to determine two critical growth stages of wheat: heading and flowering | |
| Kaur et al. | Computer vision-based tomato grading and sorting | |
| Bhatt et al. | An analysis of the performance of Artificial Neural Network technique for apple classification | |
| CN101738998B (en) | Industrial process monitoring system and method based on local discriminant analysis | |
| Zheng et al. | Identifying strawberry appearance quality by vision transformers and support vector machine | |
| Jana et al. | Detection of rotten fruits and vegetables using deep learning | |
| Alzu’bi et al. | Dates fruits classification using SVM | |
| Arwatchananukul et al. | Fast and Efficient Cavendish Banana Grade Classification using Random Forest Classifier with Synthetic Minority Oversampling Technique. | |
| Kaushik et al. | Comparative Analysis of Cassava Leaf Disease Prediction Using the Deep Learning Approach | |
| Rahman et al. | Deep learning modeling for potato breed recognition | |
| Jana et al. | Classification of aromatic and non-aromatic rice using electronic nose and artificial neural network | |
| Aparna et al. | Detection of weed using visual attention model and SVM classifier | |
| CN109444360A (en) | Juice storage phase detection algorithm based on cell neural network and electronic nose feature extraction | |
| EP2383680A1 (en) | Classification of objects in harvesting applications | |
| Ghosh et al. | Benchmarking ML and DL Models for Mango Leaf Disease Detection: A Comparative Analysis | |
| Kazerouni et al. | Comparison of modern description methods for the recognition of 32 plant species | |
| Amari et al. | Electronic nose for anchovy freshness monitoring based on sensor array and pattern recognition methods: principal components analysis, linear discriminant analysis and support vector machine | |
| Amin et al. | SAPS: Automatic saffron adulteration prediction systems, research issues, and prospective solutions | |
| Ramya et al. | A review of different classification techniques in machine learning using WEKA for plant disease detection | |
| Sklab et al. | Identification of Non-Plant Elements in Herbarium Images Using YOLO | |
| Tugrul et al. | Deep Learning‐Based Classification of Image Data Sets Containing 111 Different Seeds | |
| Biswas et al. | Feature ranking based carrot disease recognition using MIFS method | |
| Hi et al. | A deep learning approach for lantana camara weed detection and localization in the natural environment | |
| Lin et al. | Deep convolutional neural network for automatic discrimination between Fragaria× Ananassa flowers and other similar white wild flowers in fields | |
| Auleria et al. | A review on KN earest neighbour based classification for object recognition |
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 |