CN109740666B - Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD - Google Patents
Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD Download PDFInfo
- Publication number
- CN109740666B CN109740666B CN201811644498.XA CN201811644498A CN109740666B CN 109740666 B CN109740666 B CN 109740666B CN 201811644498 A CN201811644498 A CN 201811644498A CN 109740666 B CN109740666 B CN 109740666B
- Authority
- CN
- China
- Prior art keywords
- lksvd
- citrus juice
- electronic nose
- juice aroma
- dictionary
- 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.)
- Active
Links
Images
Classifications
-
- 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
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Investigating Or Analyzing Materials By The Use Of Fluid Adsorption Or Reactions (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to an electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD, which is characterized by comprising the following steps of: the method comprises the following steps: s1, collecting fragrance characteristic data of the citrus juice by using an electronic nose system; s2, filtering the citrus juice aroma characteristic data by using a random binary number to form a citrus juice aroma characteristic dictionary; s3, mapping the nonlinear data into linear data by using a kernel function; s4, initializing an LKSVD target function; s5, carrying out weighting coefficient standardization on the LKSVD target function; s6, updating and optimizing the standardized LKSVD objective function; s7, combining the optimized LKSVD objective function, and obtaining a class label of the citrus juice aroma feature dictionary by using a linear classifier; and S8, classifying the citrus juice aroma characteristic data according to the class labels of the citrus juice aroma characteristic dictionary. The invention introduces kernel function, which makes LKSVD algorithm more ideal for data processing; and the EQPSO algorithm is used for optimizing the weighting parameters of the LKSVD target function, so that a good classification effect is achieved.
Description
Technical Field
The invention relates to the technical field of gas characteristic detection, in particular to an electronic nose characteristic extraction and classification method for citrus juice aroma detection based on LKSVD.
Background
The aroma characteristics of the citrus processed products can be used as standards for food safety inspection, so that the research on the aroma characteristics of the citrus processed products is of great significance. Some overdue fruit juice beverages are illegally recycled and sold by enterprises, and whether the beverages are outdated or not is judged by analyzing the aroma characteristics of the fruit juice, so that the method is helpful for consumers and facilitates the work of supervision departments. Flavor is a key factor in determining citrus quality and nutritional value. At present, the main methods for detecting the quality of oranges are sensory analysis and precise instruments.
Sensory analysis is widely used in a number of conventional processes. Although this test method is very consumer friendly, there are too many subjective factors in the process, which can lead to inaccurate test results. The precision instrument analysis method which is objective and scientific in analysis becomes a main analysis test method for scientific research. However, the cost of the high-precision instrument is high, the high-precision instrument is difficult to apply in real production and life, the pretreatment step of the analysis of the high-precision instrument is complex, the operation is extremely inconvenient, and the technical requirement on operators is high, so the high-precision instrument is difficult to widely apply.
At present, an electronic nose (E-nose) which is a device designed by a similar biological olfaction device and used for analyzing gas/smell is provided, wherein an electronic nose system combines a series of sensors and an LKSVD algorithm, firstly, corresponding curves of the sensors are extracted, and characteristic data extracted by each sensor is processed by the LKSVD algorithm. However, the current electronic nose system has the following defects:
(1) the processing of nonlinear data is not ideal;
(2) the cross-responsiveness of the multiple sensors makes the acquired data redundant;
(3) the accuracy of the final classification result is affected by the relative size of the weight coefficient of the target function in the LKSVD algorithm, and the optimal weight is not obtained at present.
Disclosure of Invention
Aiming at the defects of the electronic nose system, the invention provides the electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD, and the method can obtain an accurate electronic nose feature classification label, so that the gas feature classification result acquired by the electronic nose system is more ideal.
In order to achieve the purpose, the technical scheme of the invention is as follows:
an electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD is characterized in that: the method comprises the following steps:
s1, collecting fragrance characteristic data of the citrus juice by using an electronic nose system to obtain an electronic nose characteristic matrix;
s2, filtering the citrus juice aroma characteristic data by using a random binary number, selecting the optimal combination of the sensors in the electronic nose, and forming a citrus juice aroma characteristic dictionary from all data acquired by the optimal combination;
s3, mapping the nonlinear data in the citrus juice aroma characteristic dictionary into linear data by utilizing a kernel function;
s4, initializing an LKSVD target function;
s5, carrying out weighting coefficient standardization on the LKSVD target function;
s6, updating and optimizing the LKSVD target function standardized in the step S5;
s7, combining the LKSVD objective function optimized in the step S6, and obtaining a class label of the orange juice aroma feature dictionary by using a linear classifier;
and S8, classifying the citrus juice aroma characteristic data according to the class labels of the citrus juice aroma characteristic dictionary.
By adopting the scheme, the LKSVD algorithm is not ideal in effect when being used for directly processing data because the distribution of the electronic nose data is nonlinear, and the nonlinear data can be mapped into linear data by introducing the kernel function and combining the kernel function with the LKSVD algorithm, so that the LKSVD algorithm can better process the electronic nose data; for the redundancy existing in the data acquired by the electronic nose, the random binary number is used for filtering the characteristic data, so that the redundancy can be well removed; and then, the weighting coefficient of the LKSVD objective function is standardized to obtain an accurate weight value, so that the classification effect becomes very ideal.
The selection of the appropriate dictionary is the first step, and most important, of the class-based sparse representation. The goal of dictionary learning is to learn an overcomplete dictionary matrix D: d is belonged to R n×K ， n<And K, finding a linear representation in a dictionary containing K vectors to represent the current n-dimensional vector. Signal vector y (y ∈ R) n ) Can be sparsely represented by a linear combination of atoms, denoted as x, x ∈ R K Satisfy y ═ Dx or y ≈ Dx, x can pass through | | | y-Dx | | count p Some small values of ≦ ε and some L p Norm (eg: L) 1 ，L 2 Or L ∞ Norm) to be refined. If D is a full rank matrix, the sparse representation has infinite solutions. To ensure sparsity, it is preferable to choose to include non-zero coefficientsThe solution with the least number realizes D by minimizing reconstruction errors and satisfying sparse constraints.
Further describing, in step S2, the random binary number is obtained by EQPSO algorithm.
With the above scheme, the proper dictionary initialization atoms are filtered by using random binary numbers, wherein 1 in the binary numbers indicates that the sensor is selected, and 0 indicates that the sensor is not selected, so that the effects of removing redundant information and filtering characteristic representative data are achieved. By using the recognition rate as a fitness function, only if the corresponding fitness function reaches a maximum can a set of parameters be selected to decide which sensors to use.
Described further, in step S3, the kernel function is an RBF kernel function, and its expression is:
wherein Z is i ,Z j For any two input citrus juice aroma characteristic data, sigma is a width parameter of the function, and the radial action range of the function is controlled.
In the above scheme, the RBF kernel function is a gaussian kernel function, σ determines data distribution mapped to a high-dimensional space, which is a very important parameter, and the optimal value is determined by using an EQPSO optimization algorithm.
To be further described, in step S4, the LKSVD target function expression is:
wherein the content of the first and second substances,
which is indicative of the error of the reconstruction,
representing a discriminant sparse coding error,
representing classification errors, wherein epsilon is a weighting coefficient of a discriminant sparse coding error, and omega is a weighting coefficient of the classification errors;
Y＝[y 1 ...y N ]∈R n×N is an N-dimensional input signal, X is a sparse representation of Y, and X ═ X 1 ,...x N ]∈R K ×N ，D＝[d 1 ...d K ]∈R n×K Is a dictionary for learning, K represents the number of dictionary entries, | | x i || 0 Is x i The number of non-zero terms in the vector, T is a sparse constraint factor; q is a discriminative sparse representation of the input signal Y used for classification,
h is the class label of the input signal Y, [ H ] 1 ...h N ]∈R m×N ，h i ＝[0,0...1...0,0] t ∈R m Is the label vector corresponding to the input signal, a is the linear transformation matrix, and W is the classifier parameters.
In the scheme, the KSVD algorithm is a label consistency KSVD algorithm for learning sparse coding identification dictionaries, and the algorithm jointly learns a single overcomplete dictionary and an optimal linear classification generation dictionary, so that feature points with the same class of labels have similar sparse codes. If q is i Is present in the input signal and dictionary term d k At indexes sharing the same label, then call q i Is discriminant sparse coding corresponding to the input signal Y; h is i The non-zero position in (b) indicates the class of the input signal.
Further, the method for initializing the LKSVD target function includes:
based on sparsityTo represent
Calculating to obtain parameters D0 and X 0 (ii) a By using
Calculate A 0 By using
Calculate W 0 。
In the above scheme, initialization D 0 I.e., several KSVD iterations are taken in each class, all the outputs of each KSVD are combined, and then the dictionary entry d is based on k The corresponding class initializes its labels and remains fixed throughout the dictionary learning process.
Further, in step S5, the weighting factor normalization method of the LKSVD objective function includes:
s51, re-weighting the LKSVD target function to obtain an expression:
wherein, α, β and γ are weighting factors, α + β + γ ═ 1, α, β, γ ∈ (0, 1);
and S52, determining the optimal values of alpha, beta and gamma by adopting an EQPSO algorithm.
Particle Swarm Optimization (PSO) is an evolutionary algorithm based on the group intelligence theory. The group intelligence algorithm for the continuous search space problem is widely applied due to simple programming and high convergence speed. The expression is as follows:
and is
Where t is the current iteration number of the algorithm, r1 and r2 are [0, 1 ]]The random number of (1). P id Is the current optimum position, P, of the particle gd Is the global optimum position of the particle swarm. However, in the classical particle swarm algorithm, the particle search process is implemented in an orbital fashion, and the flight speed of the particles is limited. Thus, the particle search space is limited to cover the entire feasible search space during the search process. The general PSO algorithm cannot guarantee convergence to a globally optimal solution with a probability of 1. To solve this drawback of the particle swarm algorithm, a quantum-behaved particle swarm algorithm (QPSO) is proposed.
QPSO combines particle swarm optimization algorithms with quantum mechanics. The QPSO algorithm has great advantages in the aspects of searching capability, convergence speed, precision, solving robustness and the like. Compared with other algorithms, one of the biggest characteristics of QPSO is simple calculation and few control parameters. The algorithm is not only superior to a common particle swarm optimization algorithm in search capability, but also superior to the common particle swarm optimization algorithm in precision.
Using wave functions
To determine the state of each particle and to determine the center average optimum position of the optimum positions of each particle. For the position of the particle, an updated equation is obtained:
u ij (t)～U(0,1)
where α is a compression expansion factor used to adjust the convergence rate of the particles. However, when the number of iterations is not infinite, QPSO cannot guarantee that a global optimum is found.
Enhanced QPSO (eqpso) is a modified QPSO algorithm that can guarantee that the value closest to the optimal value is found with a limited number of iterations. The algorithm has stronger particle diversity in the initial iteration stage and better local search capability in the later iteration stage. Meanwhile, the convergence rate of the EQPSO algorithm is faster than that of the PSO algorithm and the QPSO algorithm.
The EQPSO algorithm expression is as follows:
where U (0,1) is a random number distributed between 0 and 1, P id Is a local attractor, i.e. the d-dimensional component of the individual best position pbest of particle i, P gd Is the global attractor, i.e., the d-th component, L, of the group best position, mbest c Is the total number of iterations, C c Is the current iteration number.
In the scheme, the EQPSO algorithm is an improved quantum behavior particle swarm algorithm, and the EQPSO algorithm is optimized to alpha 1 ，β 1 And gamma 1 Can be described as the following steps: first, the particles are initialized to X id ＝P id ，(X id Iterating the d-dimensional component of the particle i-position vector), and then updating p according to the fitness function id And updates p at the optimal position of each particle gd Then calculate mbest, P id 、 X id Repeating the steps until the algorithm meets the end condition, and finally obtaining more accurate and optimized coefficients alpha, beta and gamma; wherein, the mbest is the average optimal position of all particles, and the calculation method is as follows:
to be further described, in step S6, the method for updating and optimizing the dictionary includes:
S61、is provided with
Where t is the current iteration number, Y new Is the value of Y for the current number of iterations, D new And D is the value of the current iteration number, then:
s62, judging D in the current iteration times new If the value is less than the threshold value, the step S63 is executed, otherwise, the value of t is added with 1, and the step S61 is executed again;
s63, obtaining D, A and W after updating through the above formula.
In the above scheme, the threshold is an empirical value and can be set by a worker according to actual conditions. Because the initial dictionary is often not optimal, and the sparsely represented data and the original data have large errors, the optimization needs to be updated row by row and column by column under the condition of meeting the sparsity, the overall error is reduced, and the available dictionary is approximated.
In step S7, the method for calculating the class label includes:
calculating y i Is sparse representation of
Calculating y using a linear classifier f (x; W) ═ Wx i Class label of
l is a class label vector;
wherein D, A and W are changed to D, A and W due to L2 norm normalization in the algorithm
Form, applied to a linear classifier.
Has the advantages that: (1) filtering data acquired by an electronic nose system by using random binary number, and selecting an optimal sensor combination to remove redundant information generated by cross response of a plurality of sensors; (2) a kernel function is introduced, and nonlinear data acquired by an electronic nose system is mapped into linear data, so that the LKSVD algorithm is more ideal in data processing; (3) and optimizing the weighting parameters of the LKSVD target function by using an EQPSO algorithm to obtain the optimal weight, thereby achieving a good classification effect.
Drawings
FIG. 1 is a block flow diagram of the present invention;
FIG. 2 is a detailed flow chart of the algorithm of the present invention.
Detailed Description
The invention is further illustrated by the following examples and figures:
example (b):
in this embodiment, the experimental room temperature (25 ℃) and the humidity environment (40%) are set, so that each sensor of the electronic nose system can normally work. Placing valencia orange juice in an electronic nose system and sampling gas data in three stages, first, exposing all sensors to clean air for 5 minutes to obtain a baseline; secondly, the target gas was introduced into the chamber for 7 minutes; third, the sensor array was again exposed to clean air for 5 minutes to clean his test and return to baseline. And finally, extracting the maximum value of the steady-state response of the sensor and establishing a characteristic matrix of the electronic nose.
Entering a data processing step: as shown in fig. 1 and 2, the electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD is characterized in that: the method comprises the following steps:
s1, collecting fragrance characteristic data of the citrus juice by using an electronic nose system to obtain an electronic nose characteristic matrix;
s2, filtering the citrus juice aroma characteristic data by using a random binary number, selecting the optimal combination of the sensors in the electronic nose, and forming a citrus juice aroma characteristic dictionary from all data acquired by the optimal combination; the random binary number is obtained by an EQPSO algorithm;
s3, mapping the nonlinear data in the citrus juice aroma characteristic dictionary into linear data by utilizing a kernel function; the kernel function is an RBF kernel function, and the expression is as follows:
wherein Z is i ,Z j For any two input citrus juice aroma characteristic data, sigma is a width parameter of the function, and the radial action range of the function is controlled;
s4, initializing an LKSVD target function;
the LKSVD target function expression is as follows:
wherein, the first and the second end of the pipe are connected with each other,
which is indicative of the error of the reconstruction,
representing a discriminant sparse coding error,
representing classification errors, wherein epsilon is a weighting coefficient of discriminant sparse coding errors, and omega is a weighting coefficient of the classification errors;
Y＝[y 1 …y N ]∈R n×N is an N-dimensional input signal, X is a sparse representation of Y, and X ═ X 1 ,...x N ]∈R K×N ，D＝[d 1 ...d K ]∈R n×K Is a learning dictionary, K represents the number of dictionary entries, | | x i || 0 Is x i The number of non-zero terms in the vector, T is a sparse constraint factor; q is a discriminative sparse representation of the input signal Y used for classification,
h is the class label of the input signal Y, [ H ] 1 ...h N ]∈R m×N ，h i ＝[0,0...1...0,0] t ∈R m Is a label vector corresponding to the input signal, a is a linear transformation matrix, W is a classifier parameter;
the initialization method of the LKSVD target function comprises the following steps:
sparse representation based
Calculating to obtain a parameter D 0 And X 0 (ii) a By using
Calculate A 0 By using
Calculate W 0
S5, carrying out weighting coefficient standardization on the LKSVD target function;
the weighting coefficient standardization method of the LKSVD objective function comprises the following steps:
s51, re-weighting the LKSVD target function to obtain an expression:
wherein, α, β and γ are weighting factors, α + β + γ ═ 1, α, β, γ ∈ (0, 1);
and S52, determining the optimal values of alpha, beta and gamma by adopting an EQPSO algorithm.
In this embodiment, α is 0.1988, β is 0.7661, and γ is 0.0351
The EQPSO algorithm expression is as follows:
where U (0,1) is a random number distributed between 0 and 1, P id Is a local attractor, i.e. the d-dimensional component of the individual best position pbest of particle i, P gd Is the global attractor, i.e., the d-th component, L, of the group best location, mbest c Is the total number of iterations, C c Is the current iteration number.
S6, updating and optimizing the LKSVD target function standardized in the step S5; the updating optimization method of the dictionary comprises the following steps:
s61, setting
Where t is the number of update iterations, Y new Is the value of Y for the current number of iterations, D new And D is the value of the current iteration number, then:
s62, judging D in the current iteration times new If the value is less than the threshold value, the step S63 is executed, otherwise, the value of t is added with 1, and the step S61 is executed again;
s63, obtaining D, A and W after updating through the above formula.
S7, combining the LKSVD objective function optimized in the step S6, and obtaining a class label of the citrus juice aroma feature dictionary by using a linear classifier;
the class label calculation method comprises the following steps:
calculating y i Is sparse representation of
Calculating y using a linear classifier f (x; W) ═ Wx i Class label of
l is a class label vector;
wherein D, A and W are changed to D, A and W due to L2 norm normalization in the algorithm
Form, applied to a linear classifier.
And S8, classifying the citrus juice aroma characteristic data according to the class labels of the citrus juice aroma characteristic dictionary.
In this example, several different KSVD techniques were compared and the results are given in the following table:
the classification accuracy of the LKSVD algorithm can be improved by introducing the kernel function.
The information loss can be caused by the excessively small fragrance characteristic dictionary of the citrus juice, so that the recognition rate is low; when the dictionary is too large, the process is complicated and the recognition rate is not necessarily high. When different sensors are selected to initialize the citrus juice aroma feature dictionary, the results are also different. The results of the initialization technique for different citrus juice aroma feature dictionaries are compared as follows:
Claims (9)
1. an electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD is characterized in that: the method comprises the following steps:
s1, collecting fragrance characteristic data of the citrus juice by using an electronic nose system to obtain an electronic nose characteristic matrix;
s2, filtering the citrus juice aroma characteristic data by using a random binary number, selecting the optimal combination of the sensors in the electronic nose, and forming a citrus juice aroma characteristic dictionary from all data acquired by the optimal combination;
s3, mapping the nonlinear data in the citrus juice aroma characteristic dictionary into linear data by utilizing a kernel function;
s4, initializing an LKSVD target function;
s5, carrying out weighting coefficient standardization on the LKSVD target function;
s6, updating and optimizing the LKSVD target function standardized in the step S5;
s7, combining the LKSVD objective function optimized in the step S6, and obtaining a class label of the citrus juice aroma feature dictionary by using a linear classifier;
and S8, classifying the citrus juice aroma characteristic data according to the class labels of the citrus juice aroma characteristic dictionary.
2. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 1, wherein: in step S2, the random binary number is obtained by the EQPSO algorithm.
3. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 1, wherein: in step S3, the kernel function is an RBF kernel function, and its expression is:
wherein Z is i ,Z j For any two input citrus juice aroma characteristic data, sigma is a width parameter of the function,the radial extent of action of the control function.
4. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 1, wherein: in step S4, the LKSVD objective function expression is:
wherein the content of the first and second substances,
which is indicative of the error of the reconstruction,
representing a discriminant sparse coding error,
representing classification errors, wherein epsilon is a weighting coefficient of discriminant sparse coding errors, and omega is a weighting coefficient of the classification errors;
Y＝[y 1 ...y N ]∈R n×N is an N-dimensional input signal, X is a sparse representation of Y, and X is [ X ] 1 ,...x N ]∈R K×N ，D＝[d 1 ...d K ]∈R n×K Is a dictionary for learning, K represents the number of dictionary entries, | | x i || 0 Is x i Number of non-zero entries in the vector, where i ∈ [1, N ]]T is a sparse constraint factor; q is a discriminative sparse representation of the input signal Y used for classification,
h is the class label of the input signal Y, [ H ] 1 ...h N ]∈R m×N ，h i ＝[0,0...1...0,0] t ∈R m Is the label vector corresponding to the input signal, a is the linear transformation matrix, and W is the classifier parameter.
5. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 4, wherein: the initialization method of the LKSVD target function comprises the following steps:
sparse representation based
Calculating to obtain a parameter D 0 And X 0 (ii) a By using
Calculate A 0 By using
Calculate W 0 。
6. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 5, wherein: in step S5, the weighting factor normalization method for the LKSVD objective function includes:
s51, re-weighting the LKSVD target function to obtain an expression:
wherein, α, β and γ are weighting factors, α + β + γ ═ 1, α, β, γ ∈ (0, 1);
and S52, determining the optimal values of alpha, beta and gamma by adopting an EQPSO algorithm.
7. An LKSVD-based electronic nose feature extraction and classification method for citrus juice aroma detection according to claim 2 or 6, characterized in that: the EQPSO algorithm expression is as follows:
where U (0,1) is a random number distributed between 0 and 1, P id Is a local attractor, i.e. the d-dimensional component of the individual best position pbest of particle i, P gd Is the global attractor, i.e., the d-th dimension component, L, of the group best location, mbest c Is the total number of iterations, C c Is the current iteration number.
8. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 6, wherein: in step S6, the updating optimization method of the dictionary includes:
s61, setting
Where t is the number of update iterations, Y new Is the value of Y for the current number of iterations, D new And D is the value of the current iteration number, then:
s62, judging D in the current iteration times new If the value is less than the threshold value, the step S63 is executed, otherwise, the value of t is added with 1, and the step S61 is executed again;
s63, obtaining D, A and W after updating through the above formula.
9. The LKSVD-based citrus juice aroma detection-oriented electronic nose feature extraction and classification method according to claim 8, wherein: the class label calculation method comprises the following steps:
calculating y i Is sparse representation of
Calculating y using a linear classifier f (x; W) ═ Wx i Class label of
l is a class label vector;
wherein D, A and W are changed into the norm of L2 in the algorithm for normalization
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811644498.XA CN109740666B (en) | 2018-12-29 | 2018-12-29 | Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811644498.XA CN109740666B (en) | 2018-12-29 | 2018-12-29 | Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109740666A CN109740666A (en) | 2019-05-10 |
| CN109740666B true CN109740666B (en) | 2022-08-26 |
Family
ID=66362777
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811644498.XA Active CN109740666B (en) | 2018-12-29 | 2018-12-29 | Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740666B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112433028B (en) * | 2020-11-09 | 2021-08-17 | 西南大学 | Electronic nose gas classification method based on memristor cell neural network |
| CN118010935A (en) * | 2024-03-11 | 2024-05-10 | 广东派道嗅辨科技有限公司 | Method and device for sniffing automobiles |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104504407A (en) * | 2014-12-17 | 2015-04-08 | 西南大学 | Electronic nose feature selection optimization method on basis of multiple Fisher kernel discriminant analysis |
-
2018
- 2018-12-29 CN CN201811644498.XA patent/CN109740666B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104504407A (en) * | 2014-12-17 | 2015-04-08 | 西南大学 | Electronic nose feature selection optimization method on basis of multiple Fisher kernel discriminant analysis |
Non-Patent Citations (2)
| Title |
|---|
| DKSVD和SVM在钢轨裂纹漏磁识别上的对比研究;张陆唯等;《电子测量技术》;20170915(第09期);200-203 * |
| 基于机器嗅觉的柑橘品种无损检测与识别;彭珂等;《江西农业大学学报》;20171020(第05期);181-188 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109740666A (en) | 2019-05-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Berrueta et al. | Supervised pattern recognition in food analysis | |
| CN111699494A (en) | Object recognition using convolutional neural networks trained by principal component analysis and repetitive spectral clustering | |
| CN111259140B (en) | False comment detection method based on LSTM multi-entity feature fusion | |
| AU2009241650B2 (en) | Image processing | |
| CN107203787A (en) | Unsupervised regularization matrix decomposition feature selection method | |
| CN109740666B (en) | Electronic nose feature extraction and classification method for citrus juice aroma detection based on LKSVD | |
| CN111340132B (en) | Machine olfaction mode identification method based on DA-SVM | |
| CN112905739B (en) | False comment detection model training method, detection method and electronic equipment | |
| CN111046961B (en) | Fault classification method based on bidirectional long-time and short-time memory unit and capsule network | |
| CN109299246B (en) | Text classification method and device | |
| CN107480441B (en) | Modeling method and system for children septic shock prognosis prediction | |
| CN109145685B (en) | Fruit and vegetable hyperspectral quality detection method based on ensemble learning | |
| Tang et al. | An MLP network based on residual learning for Rice hyperspectral data classification | |
| Vidyarthi et al. | Classification of first quality fancy cashew kernels using four deep convolutional neural network models | |
| CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
| Çakmak | Grapevine Leaves Classification Using Transfer Learning and Fine Tuning | |
| Abas et al. | Agarwood oil quality classifier using machine learning | |
| Huang et al. | Resolving intra-class imbalance for gan-based image augmentation | |
| Cudic et al. | Prediction of sorghum bicolor genotype from in-situ images using autoencoder-identified SNPs | |
| CN114036947A (en) | Small sample text classification method and system for semi-supervised learning | |
| CN109444360B (en) | Fruit juice storage period detection algorithm based on cellular neural network and electronic nose feature extraction | |
| Khazaei et al. | Evaluation of statistical and neural network architectures for the classification of paddy kernels using morphological features | |
| Wu et al. | Identification of early moldy rice samples by PCA and PNN | |
| Li et al. | Genetic algorithms (GAs) and evolutionary strategy to optimize electronic nose sensor selection | |
| Hancock et al. | Supervised hierarchical clustering using CART |
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 |