Electronic nose array optimization method based on fuzzy rough set and mutual information theory

Abstract

The invention relates to the technical field of sensors, in particular to an electronic nose array optimization method based on a fuzzy rough set and a mutual information theory, a hierarchical screening strategy is adopted, a sensor array is divided into three layers, the first layer is a base layer sensor which is obtained through screening according to the fuzzy rough set theory and has great contribution to label classification; the second layer is an enhancement layer sensor which is obtained through screening according to a mutual information theory and has relatively large association with the classification label; the third layer is a comprehensive layer sensor which is obtained through screening and has relatively high internal attribute importance degree and relatively small redundant mutual information; the comprehensive layer sensor set is composed of several sensors with the strongest gas recognition capacity in the gas recognition process, namely optimal sensors; by means of the electronic nose array optimization method, in the gas recognition process, the electronic nose system can use few optimal sensors to achieve high recognition precision, and the complexity and cost of system design are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of sensors, and particularly to an optimization method for an electronic nose array based on fuzzy rough sets and mutual information theory. Background Art

[0002] An electronic nose mainly consists of a gas sensor array, signal preprocessing, and pattern recognition. As the core component of an electronic nose system, the gas sensor array often causes problems such as selective overlap, high computational cost, low detection efficiency, and data redundancy due to its strong cross-sensitivity, resulting in difficulty in further improving the performance of the electronic nose system. Array optimization can minimize the number of sensors to the greatest extent while ensuring detection accuracy, thereby reducing the scale of the gas sensor array and lowering the system cost. The methods for optimizing an electronic nose array mainly include search-based methods, variable selection methods, neural network-based methods, and feature selection methods.

[0003] In the prior art, the main means of feature selection algorithms is to evaluate the redundancy and correlation of features through fuzzy rough sets or information theory; however, there are usually problems such as incomplete elimination of redundant features and loss of features indispensable for label classification, making it difficult to achieve the optimal combination of feature subsets.

[0004] Therefore, there is an urgent need for an optimization method for an electronic nose array that can more thoroughly eliminate redundant features and effectively retain features related to label classification. Summary of the Invention

[0005] In view of this, the present invention discloses an optimization method for an electronic nose array based on fuzzy rough sets and mutual information theory to solve the above problems; the method includes the steps of:

[0006] S1. Obtain classification labels, time response sequences, and a sample set, and establish an original feature set;

[0007] S2. Based on fuzzy rough set theory, calculate the granularity of each sensor feature in the original feature set according to the sample set;

[0008] S3. According to the granularity of the sensor features, perform a preliminary screening on the original feature set to obtain a basic layer sensor set;

[0009] S4. According to mutual information theory, calculate the correlation between each sensor in the basic layer sensor set and the classification label;

[0010] S5. According to the magnitude of the correlation, perform a secondary screening on the basic layer sensor set to obtain an enhanced layer sensor set;

[0011] S6. Calculate the internal attribute importance of the corresponding features of the sensors in the enhanced layer sensor set and the redundant mutual information between sensors, and obtain the importance index of the features by combining the internal attribute importance and the redundant mutual information;

[0012] S7. Remove the sensors in the enhanced layer sensor set according to the importance index, and the remaining sensors form the comprehensive layer sensor set.

[0013] The beneficial effects of the present invention include:

[0014] The present invention effectively solves the problems of high redundancy in sensors and easy removal of important sensors by fully combining the fuzzy rough set theory and the mutual information theory to optimize the sensor array of the electronic nose; through the hierarchical screening strategy, specific evaluation indexes are concerned hierarchically during the sensor selection process, thereby effectively improving the accuracy of sensor selection and avoiding deviation problems that may be caused by a single standard; through the principle of evaluating importance by the attribute importance of concentrated features, the correlation and redundancy in mutual information, hierarchical screening is realized,

[0015] The step-by-step optimization strategy realizes the optimization of the electronic nose array, providing a new solution for the design field of the electronic nose. Description of the Drawings

[0016] Figure 1 It is the flow structure diagram of the embodiment of the present invention;

[0017] Figure 2 It is the Venn diagram showing the correlation between the sensors in the basic layer sensor set and the classification labels of the present invention;

[0018] Figure 3 It is the Venn diagram showing the redundancy between the sensors in the enhanced layer sensor set of the present invention. Detailed Embodiment

[0019] In order to make the purpose, technical solution, features and advantages of the present invention clearer and more understandable, the present invention will be further described below with reference to the accompanying drawings and embodiments. Figure 1 The flow structure of this embodiment is shown. This embodiment includes the steps:

[0020] S1. Obtain the classification labels, time response sequences, sample sets, and establish the original feature set.

[0021] Specifically, the classification label refers to the label corresponding to a certain gas during odor recognition. For example, when identifying a mixed gas, the classification labels include ethanol, formaldehyde, etc. The classification label data set of this embodiment uses the gas sensor array under flow modulation of UCI, which is mainly used to identify acetone, ethanol and their mixtures; in this embodiment, the sensor array includes 16 sensors.

[0022] The input original feature set refers to the maximum value of the response of the sensor to the target gas within a certain time series. The original feature set F = {X 1 , X 2 , …, X n}, and the classification label Y = {Y 1 , Y 2 , …, Y l}. Among them, the original feature set F is the maximum value corresponding to the sensor under the target gas within a certain time series range; n represents the number of sensors, X i represents the maximum value of the response of the i-th sensor within a certain time range, the label variable Y is the type of the target gas sample, and l represents the number of classification label samples. The original sensor array refers to a combination of sensors that has not been processed or optimized. The sensor X i is the current sensor in the current sensor set.

[0023] S2. Based on the fuzzy rough set theory, calculate the granularity of each sensor feature in the original feature set according to the sample set.

[0024] Specifically, traverse the original feature set F and calculate the granularity GP U (X i ). Before calculating the granularity of each feature, use the corresponding sensor X i to divide the equivalent classes corresponding to the sensor in the sample set U, and define GP U (X i ) as:

[0025]

[0026] Among them, X i is the currently evaluated sensor, c is the number of equivalent classes divided by the sensor X i , |X i | is the number of sensors equivalent to X i , and |U| is the number of the sample set.

[0027] Further, among them, c is the number of equivalent classes divided by the sensor X i , and the equivalent class is the class element that conforms to the equivalence relationship. The equivalence relationship refers to defining a binary relationship R on the sample set U and satisfying the following three properties:

[0028] Reflexivity:

[0029] Symmetry:

[0030] Transitivity:

[0031] Among them, an equivalence relation R on the sample set U is defined. For an element x in the set U, the subset composed of all elements in the set U that have an equivalence relation with the element x is called the equivalence class corresponding to x.

[0032] S3. According to the granularity of the sensor features, the original feature set is preliminarily screened to obtain the basic layer sensor set.

[0033] Specifically, the sensors with a sensor feature granularity greater than the first threshold are added to the basic layer sensor set xs. The first threshold is preferably the mean value of all sensor feature granularities in the original feature set, so as to facilitate the screening of sensors that contribute more than the average level to the data. The basic layer sensor set is represented as:

[0034] xs = xs ∪ X i

[0035] where i ≤ n, n is the number of sensors in the original feature set, xs is the basic layer sensor set, and it is an empty set initially.

[0036] S4. According to the mutual information theory, calculate the correlation between each sensor in the basic layer sensor set and the classification label.

[0037] Specifically, traverse the basic layer sensor set xs, and calculate the correlation magnitude between each sensor in the subset xs and the label Y one by one. Use the mutual information I(X i ; Y) between the features of the sensor and the label as an index to evaluate the correlation magnitude between this sensor and the target gas. Define R(X i ):

[0038] R(X i ) = I(X i ; Y)

[0039] where X i represents the currently evaluated sensor in the basic layer sensor set, and Y represents the classification label.

[0040] S5. According to the correlation magnitude, perform a secondary screening on the basic layer sensor set to obtain the enhanced layer sensor set.

[0041] Specifically, use the correlation magnitude to screen the sensors in the basic layer sensor set xs, and select the sensors with a correlation magnitude between the sensor and the classification label greater than the second threshold to form the enhanced layer sensor set s. The second threshold is preferably the mean value of the correlation magnitudes between all sensors and the classification label in the basic layer sensor set, so as to facilitate the screening of sensors with a correlation with the classification label higher than the average level.

[0042] Furthermore, the sensors with a correlation between the sensor classification label and the second threshold greater than the second threshold are added to the enhanced layer sensor set s, that is

[0043] s = s ∪ X i

[0044] where i ≤ m, m is the number of sensors in the basic layer sensor set xs, and s is an empty set initially.

[0045] Furthermore, as Figure 2 shown, Region ① is the mutual information between the features corresponding to sensor X i and label Y, that is, the correlation between the sensors in the basic layer sensor set and the classification label. Among them, X i is a sensor in the basic layer sensor set xs.

[0046] S6. Calculate the internal attribute importance of the features corresponding to the sensors in the enhanced layer sensor set and the redundant mutual information between sensors, and combine the internal attribute importance and the redundant mutual information to obtain the importance index of the features.

[0047] Specifically, the internal attribute importance in the fuzzy rough set theory can be used to quantify the independent contribution of features and to determine whether the sensor corresponding to the feature is irreplaceable in the current sensor set. Traverse the sensor set s and calculate the internal attribute importance of the sensor features and the magnitude of the redundant mutual information Rd(X i ); The definition of is as follows:

[0048]

[0049] where s is the sensor set, X i is the currently evaluated sensor, Y is the label, inner represents the internal attribute of the feature, U represents the total number of sample sets, and GP U represents the granularity of the feature.

[0050] Calculating the internal attribute importance of the sensor feature and the class label Y can be used to evaluate the indispensability of the current sensor for the classification label Y. The first term in the formula represents the granularity relative to label Y after removing the current sensor on the basis of the current sensor set, and the second term represents the granularity relative to label Y under the condition of the current sensor set. The difference between the two represents the importance of the current sensor X i with respect to the sensor set s relative to label Y. If the internal attribute importance is relatively high, it indicates that the sensor plays a key role in gas recognition and cannot be easily removed; otherwise, vice versa.

[0051] Evaluate the redundancy between sensors by combining mutual information. Define the redundancy as:

[0052]

[0053] where Xi represents the i-th sensor being currently evaluated, X j represents the other sensors in the sensor set s, and I represents the magnitude of mutual information. When the redundancy Pd(X i ) is large, it indicates that there is a large amount of redundant information between the current sensor X i and the other sensors in the sensor set s; conversely, when the redundancy Rd(X i ) is small, it indicates that there is almost no redundant information between the current sensor X i and the other sensors in the sensor set s.

[0054] Combining the two gives the importance index J(X i ). Define J(X i ) as:

[0055]

[0056] This index combines the comprehensive measurement of the indispensability of sensors and the redundant information between sensors in sensor selection. The main purpose is to ensure that important sensors are not wrongly removed while removing redundant sensors, so as to achieve a better sensor selection. When the internal attributes of sensor features are higher and the redundancy between sensors is smaller, the score of the sensor importance index is higher; conversely, when the importance degree of the internal attributes of sensor features is lower and the redundancy between sensors is larger, the score of the sensor importance index is lower.

[0057] S7. Remove the sensors in the enhanced layer sensor set according to the importance index, and the remaining sensors form the comprehensive layer sensor set.

[0058] Specifically, according to the index for evaluating sensor redundancy, the last n sensors with the lowest scores are removed from the sensor set s as redundant sensors. n is determined according to actual requirements and the total number of sensors. In this embodiment, n is selected as 8 (50% of the total number of sensors).

[0059] Furthermore, since sensors that contribute more to the data set and have a greater correlation with the label have been selected in the base layer and the enhanced layer, the main purpose of sensor removal in the comprehensive layer is to remove sensors with redundant information. A higher importance degree of internal attributes indicates that the sensor X i is indispensable for the classification target Y. Adding this index ensures that sensors highly important for the target task can be retained during the sensor selection process, avoiding the misdeletion of key sensors due to the redundancy of other sensors.

[0060] Furthermore, as Figure 3 shown, region ③ + ② is the feature corresponding to the sensor X i in the enhanced layer sensor set s jThe mutual information of corresponding features. Region ② represents sensor X i and sensor X j The redundancy among sensor X i , class label Y. Among them, sensor X j is one of the sensors in the enhanced layer sensor set s, and X i represents all other sensors in the enhanced layer sensor set s except sensor X. By using region ③ to represent the redundancy of sensors instead of region ③ + ②, it can effectively avoid the elimination of sensors with redundant information that makes important contributions to the classification label.

[0061] The present invention adopts a hierarchical screening strategy, dividing the sensor array into three layers. The first layer is the basic layer sensors screened according to the fuzzy rough set theory, which contribute more to label classification; the second layer is the enhanced layer sensors screened according to the mutual information theory, which have a greater association with the classification label; the third layer is the comprehensive layer sensors with a greater internal attribute importance and a smaller redundant mutual information. The comprehensive layer sensor set consists of several sensors with the strongest gas recognition ability in the gas recognition process, that is, the optimal sensors. Through the electronic nose array optimization method of the present invention, in the process of gas recognition, the electronic nose system can achieve a higher recognition accuracy with fewer optimal sensors, reducing the complexity and cost of system design.

[0062] Finally, it should be noted that the above description only presents some embodiments of the present invention. For those skilled in the art, various changes, modifications, substitutions, and deformations can be conceived without departing from the principles and spirits of the present invention. The protection scope of the present invention is defined by the appended claims and their equivalents, and the above behaviors should all be covered within the protection scope of the present invention.

Claims

1. An electronic nose array optimization method based on fuzzy rough sets and mutual information theory, characterized in that: include: S1, obtain classification labels, time response sequences, sample sets, and establish original feature sets; S2, based on the fuzzy rough set theory, the granularity of each sensor feature in the original feature set is calculated according to the sample set; S3. Preliminarily screen the original feature set according to the granularity of the sensor features to obtain the base layer sensor set; S4. Calculate the correlation between each sensor in the base layer sensor set and the classification label according to the mutual information theory; S5. Perform secondary screening on the base layer sensor set according to the correlation to obtain the enhanced layer sensor set; S6, calculating the internal attribute importance of the corresponding features of the sensors in the enhanced layer sensor set and the redundant mutual information between the sensors, and combining the internal attribute importance and the redundant mutual information to obtain the importance index of the feature; S7. Eliminate sensors in the enhanced layer sensor set according to the importance index, and the remaining sensors form the comprehensive layer sensor set.

2. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1 is characterized in that: Calculate Sensor X i Granularity GP U (X i )for: Where c is the value of the sensor X i The number of equivalence classes divided, |X i |For X i The equivalent number of sensors, |U| is the total number of sensor samples.

3. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1 is characterized in that: During the initial screening, the screening condition is that the characteristic granularity of the sensor is greater than a first threshold, wherein the value of the first threshold is the average of the characteristic granularities of all sensors in the original feature set.

4. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1, characterized in that: During the secondary screening, the screening condition is that the correlation between the sensor and the classification label is greater than a second threshold, wherein the value of the second threshold is the average value of the correlation between all sensors in the basic layer sensor set and the classification label.

5. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 4 is characterized in that: Calculate Sensor X i Correlation with the classification label R(X i )for: R(X i )=I(X i ;Y) Among them, X i represents the sensor currently evaluated in the base layer sensor set, and Y represents the classification label.

6. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1, characterized in that: Calculate internal attribute importance for: Among them, X i is the sensor currently being evaluated, s is the set of enhanced layer sensors, Y is the classification label, GP U Indicates the granularity at which features are calculated.

7. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1, characterized in that: Calculate the redundant mutual information Rd(X i )for: Among them, s is the enhanced layer sensor set, X i is the i-th sensor currently evaluated in the sensor set, X j is the jth sensor currently evaluated in the sensor set, and I represents the calculated mutual information size.

8. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1, characterized in that: The feature importance index J(X i ) is: in, Represents the importance of internal attributes, Rd(X i ) represents the redundant mutual information.

9. The electronic nose array optimization method based on fuzzy rough sets and mutual information theory according to claim 1, characterized in that: When eliminating sensors in the enhanced layer sensor set, n sensors with the lowest importance are eliminated, and the value of n is 50% of the total number of sensors in the sensor array.

Images