CN111474297B - An online drift compensation method for sensors in a bionic olfactory system - Google Patents

Abstract

The invention relates to an online drift compensation method for a sensor in a bionic olfactory system, and belongs to the technical field of sensors. The method comprises the following steps: step 1) performing source domain reconstruction according to an input sample batch number; and 2) constructing a classification model by using the reconstructed source domain and target domain samples and storing a prediction result. The method comprises the steps of utilizing output response samples of sensors in a bionic olfactory system in two successive batches, namely performing source domain reconstruction on samples of a previous batch which are predicted through a classification model and samples of an initial batch which are artificially marked, and then building a classification model through condition distribution self-adaption and manifold regularization to realize online drift compensation of the sensors in the bionic olfactory system. The gas identification model can be continuously updated along with the drift of the sensor, so that the method is more suitable for actual production and use scenes of the bionic olfactory system in a real scene, and the service life of equipment can be prolonged.

Description

An online drift compensation method for sensors in a bionic olfactory system

technical field

The invention belongs to the technical field of sensors, and relates to an online drift compensation method of a sensor in a bionic olfactory system.

Background technique

The bionic olfactory system consists of a gas sensor array, a signal preprocessing unit and a pattern recognition algorithm, which can be used for gas recognition. When the gas is passed into the system, the sensor array generates corresponding electrical signal responses according to the gas characteristics, and converts the preprocessed signals into gas recognition results through a pattern recognition algorithm.

The sensor drifts due to its own aging or gas poisoning. Drift problems have long existed with the use of bionic olfactory systems and cannot be avoided. Drift will change the output response of the sensor, which will lead to the failure of the initially constructed classification model to accurately predict the samples collected later.

In recent years, many algorithms for sensor drift compensation have been proposed, mainly divided into three categories: signal preprocessing, component correction and machine learning. Although these algorithms can achieve sensor drift compensation to a certain extent, most of them are offline methods and need to be recycled regularly. equipment to complete manual correction, which is not suitable for practical application scenarios. In addition, there is a difference in the characteristic distribution of the output response samples before and after sensor drift, and the related work of drift compensation in the past has focused on reducing the edge distribution difference, without considering the influence of the conditional distribution difference.

Therefore, how to reduce the difference in conditional distribution caused by sensor drift and achieve effective online update of the classification model has a great influence on the correctness of the gas discrimination results of the bionic olfactory system. The online drift compensation method for sensors in a bionic olfactory system disclosed in this patent can build a classification model through conditional distribution adaptation and manifold regularization, and at the same time utilize source domain reconstruction to realize online update of the model, and complete the analysis of the bionic olfactory system. Compensation for samples collected by sensors that have drifted is more reasonable in real-world usage scenarios.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide an online drift compensation method for sensors in a bionic olfactory system. The samples and the initial batch of manually labeled samples are reconstructed in the source domain, and then a classification model is built through conditional distribution adaptation and manifold regularization, so as to realize the online drift compensation of sensors in the bionic olfactory system.

To achieve the above object, the present invention provides the following technical solutions:

An online drift compensation method for a sensor in a bionic olfactory system, the method includes the following steps:

Step 1) Perform source domain reconstruction according to the input sample batch number;

Step 2) Use the reconstructed source domain and target domain samples to construct a classification model and save the prediction results.

Optionally, the step 1) includes the following steps:

Step 11) input the batch number a of the sample;

n₁为初始批次样本数，否则，D_s由当前目标域的前一批次已完成分类预测的样本集

与D₁共同构造，即 D_s＝D₁∪D_a，其中n_a为a批次样本数，当a＝1时，由于是首次模型构建，不存在上一批次的分类预测结果，故此时D_s＝D₁，重构后的源域样本数为：Step 12) Reconstruct the source domain according to the batch number a. When a=1, the source domain D _s is selected as the initial batch of labeled sample sets collected when the sensor does not drift

n ₁ is the number of samples in the initial batch, otherwise, D _s is the set of samples that have been classified and predicted by the previous batch of the current target domain

It is constructed together with D ₁ , that is, D _s = D ₁ ∪ D _a , where n _a is the number of samples in batch a. When a = 1, because it is the first model construction, there is no classification prediction result of the previous batch, so When D _s = D ₁ , the number of reconstructed source domain samples is:

Optionally, the step 2) includes the following steps:

n_t为目标域样本数；Step 21) Input unlabeled a+1 batches of target domain samples

n _t is the number of samples in the target domain;

Step 22) Use principal component analysis to reduce the dimensions of X _s and X _t to p dimensions to generate subspaces S _s and S _t , G _d×p is the set of all p-dimensional subspaces, and each subspace is regarded as a Grassmannian A point on the manifold space G _d×p , let Φ(t) be a geodesic on G _d×p , where t∈[0,1], Φ(0)=S _s and Φ(1)= S _{t is} used as the two ends of the geodesic line, z _i and z _j are the eigenvectors of x _i and x _j projected on the infinite-dimensional space, and the inner product is expressed as:

In the above formula, x _i ,x _j ∈D _s ∪D _t , and G represents the geodesic kernel:

In the above formula, R _s is composed of the remaining dp-dimensional features after the p-dimensional features are extracted from X _s by PCA. U ₁ and U ₂ are mutually orthogonal matrices, which are obtained by singular value decomposition. Λ ₁ , Λ ₂ and Λ ₃ are Diagonal matrix, the element values in the matrix are:

和

表示，其中

The sample feature space obtained by projecting X _s and X _t is used

and

said, of which

Step 23) Use Z _s to train a classifier through the k-nearest neighbor algorithm with k=1, and then bring Z _t into the classifier to obtain pseudo-labels

Step 24) Use Z _s and Z _t to select a Gaussian kernel to construct the kernel function

Step 25) Determine the neighbor relationship of each point in Z _s and Z _t through the k-nearest neighbor algorithm to obtain the similarity matrix W:

In the above formula, r(z _i , z _j )=1 indicates that _zi and z _j are neighbors to each other; after obtaining W, the Laplacian matrix L is calculated:

计算获得，此时流形正则约束项表示为：In the above formula, D is a diagonal matrix, by

The calculation is obtained, and the manifold regular constraint term is expressed as:

Step 26) Conditional distribution adaptation is performed on Z _s and Z _t , and the conditional distribution difference is measured on the regenerated Hilbert space by the maximum mean difference, and approximate statistical estimation is performed using the empirical estimation formula:

由两个对角矩阵组成，其中：In the above formula, C represents the total number of categories of labels contained in the sample,

consists of two diagonal matrices, where:

表示Z_s和Z_t中标签为l的n_l个样本的分类结果集合，f表示G_d×p下的分类预测函数：

Represents the classification result set of n _l samples with label l in Z _s and Z _t , and f represents the classification prediction function under G _{d × p} :

In the above formula, α=[α ₁ ,α ₂ ,...,α _N ] ^T is the coefficient vector, and the final calculation formula of the conditional distribution adaptive term is:

MMD ² (H _K , Q _s , Q _t )=tr(f ^T Mf)

Each element in M in the above formula is directly calculated by the following formula:

According to the principle of structural risk minimization, combined with the manifold regularization term and the conditional distribution adaptation term, the final optimization objective of the classifier f is:

In the above formula, U is the indicator matrix of the domain where the sample is located, namely:

Set the partial derivative of α in the optimization equation to 0 to solve for α:

α=(λ ₁ I+(λ ₂ M+λ ₃ L+U)K) ^-1 UY ^T

的更新，重复步骤26)e次以迭代更新M和α；Use f to complete the pair

update, repeat step 26) e times to iteratively update M and α;

并保存，等待下一批次样本输入。Step 27) Obtain the predicted label of this batch of samples

and save it, waiting for the next batch of samples to be input.

The beneficial effect of the invention is that the method can continuously update the gas identification model with the drift of the sensor, which is more in line with the actual production and usage scenarios of the bionic olfactory system in real scenarios, and can prolong the service life of the equipment.

Other advantages, objects and features of the present invention will be set forth in the description which follows, to the extent that will be apparent to those skilled in the art based on a study of the following, or may be learned from is taught in the practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the following description.

Description of drawings

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be preferably described in detail below with reference to the accompanying drawings, wherein:

Fig. 1 is a flow chart of the present invention.

Detailed ways

The embodiments of the present invention are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments are only used to illustrate the basic idea of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

Among them, the accompanying drawings are only for illustrative description, and represent only schematic diagrams, not physical drawings, and should not be construed as limitations of the present invention; in order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings will be omitted, The enlargement or reduction does not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the accompanying drawings may be omitted.

The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms "upper", "lower", "left" and "right" , "front", "rear" and other indicated orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the indicated device or element must be It has a specific orientation, is constructed and operated in a specific orientation, so the terms describing the positional relationship in the accompanying drawings are only used for exemplary illustration, and should not be construed as a limitation of the present invention. situation to understand the specific meaning of the above terms.

The present invention provides a method for online drift compensation of a sensor in a bionic olfactory system. As shown in FIG. 1 , the method includes the following steps:

Step 1) Perform source domain reconstruction according to the input sample batch number;

Further, step 1) includes the following steps:

Step 11) input the batch number a of the sample;

n₁为初始批次样本数，否则，D_s由当前目标域的前一批次已完成分类预测的样本集

与D₁共同构造，即D_s＝D₁∪D_a，其中n_a为a批次样本数，当a＝1时，由于是首次模型构建，不存在上一批次的分类预测结果，故此时D_s＝D₁，重构后的源域样本数为：Step 12) Reconstruct the source domain according to the batch number a. When a=1, D _s is selected as the initial batch of labeled samples collected when the sensor does not drift

n ₁ is the number of samples in the initial batch, otherwise, D _s is the set of samples that have been classified and predicted by the previous batch of the current target domain

It is constructed together with D ₁ , that is, D _s = D ₁ ∪ D _a , where n _a is the number of samples in batch a. When a = 1, because it is the first model construction, there is no classification prediction result of the previous batch, so When D _s = D ₁ , the number of reconstructed source domain samples is:

Step 2) use the reconstructed source domain and target domain samples to construct a classification model and save the prediction result;

Further, step 2) comprises the following steps:

n_t为目标域样本数；Step 21) Input unlabeled a+1 batches of target domain samples

n _t is the number of samples in the target domain;

Step 22) Use principal component analysis to reduce the dimensions of X _s and X _t to p dimensions to generate subspaces S _s and S _t , G _d×p is the set of all p-dimensional subspaces, and each subspace can be regarded as a gram A point on the Smanian manifold space G _d×p , let Φ(t) be a geodesic on G _d×p , where t∈[0,1], Φ(0)=S _s and Φ(1 )=S _t as the two ends of the geodesic line, z _i and z _j are the eigenvectors of x _i and x _j projected on the infinite-dimensional space, and their inner product can be expressed as:

In the above formula, x _i ,x _j ∈D _s ∪D _t , and G represents the geodesic kernel:

In the above formula, R _s is composed of the remaining dp-dimensional features after the p-dimensional features are extracted from X _s by PCA. U ₁ and U ₂ are mutually orthogonal matrices, which can be obtained by singular value decomposition, Λ ₁ , Λ ₂ and Λ ₃ is a diagonal matrix, and the element values in the matrix are:

和

表示，其中

The sample feature space obtained by projecting X _s and X _t is used

and

said, of which

Step 23) Use Z _s to train a classifier through the k-nearest neighbor algorithm with k=1, and then bring Z _t into the classifier to obtain pseudo-labels

Step 24) Use Z _s and Z _t to select a Gaussian kernel to construct the kernel function

Step 25) Determine the neighbor relationship of each point in Z _s and Z _t through the k-nearest neighbor algorithm to obtain the similarity matrix W:

In the above formula, r(z _i , z _j )=1 indicates that _zi and z _j are neighbors to each other. After obtaining W, the Laplace matrix L can be calculated:

计算获得，此时流形正则约束项可表示为：In the above formula, D is a diagonal matrix, by

The calculation is obtained, and the manifold regular constraint term can be expressed as:

Step 26) Conditional distribution adaptation is performed on Z _s and Z _t , and the conditional distribution difference is measured on the regenerated Hilbert space by the maximum mean difference, and approximate statistical estimation is performed using the empirical estimation formula:

由两个对角矩阵组成，其中：In the above formula, C represents the total number of categories of labels contained in the sample,

consists of two diagonal matrices, where:

表示Z_s和Z_t中标签为l的n_l个样本的分类结果集合，f表示G_d×p下的分类预测函数：

Represents the classification result set of n _l samples with label l in Z _s and Z _t , and f represents the classification prediction function under G _{d × p} :

In the above formula, α=[α ₁ ,α ₂ ,...,α _N ] ^T is the coefficient vector, and the final calculation formula of the conditional distribution adaptive term is:

MMD ² (H _K , Q _s , Q _t )=tr(f ^T Mf)

Each element in M in the above formula can be directly calculated by the following formula:

According to the principle of structural risk minimization, combined with the manifold regularization term and the conditional distribution adaptation term, the final optimization objective of the classifier f is:

In the above formula, U is the indicator matrix of the domain where the sample is located, namely:

Set the partial derivative of α in the optimization equation to 0 to solve for α:

α=(λ ₁ I+(λ ₂ M+λ ₃ L+U)K) ^-1 UY ^T

的更新，重复步骤26)e次以迭代更新M和α；Use f to complete the pair

update, repeat step 26) e times to iteratively update M and α;

并保存，等待下一批次样本输入。Step 27) Obtain the predicted label of this batch of samples

and save it, waiting for the next batch of samples to be input.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should all be included in the scope of the claims of the present invention.

Claims (1)

1. An on-line drift compensation method of a sensor in a bionic olfactory system is characterized in that: the method comprises the following steps:

step 1) performing source domain reconstruction according to an input sample batch number;

step 2) constructing a classification model by using the reconstructed source domain and target domain samples and storing a prediction result;

the step 1) comprises the following steps:

step 11) inputting the batch number a of the sample;

step 12) carrying out source domain reconstruction according to the batch number a, and when a is 1, D_sSelecting an initial lot of labeled sample sets collected when the sensor is not drifting

n₁Is the number of initial batch samples, otherwise, D_sSample set with class prediction completed by previous batch of current target domain

And D₁Common construction, i.e. D_s＝D₁∪D_aWherein n is_aFor the number of samples in batch a, when a is 1, there is no classification prediction result of the previous batch due to the first model construction, so D_s＝D₁The number of reconstructed source domain samples is:

the step 2) comprises the following steps:

step 21) inputting unlabeled a +1 batch target domain samples

n_tIs the number of samples in the target domain;

step 22) analysis of X using principal Components_sAnd X_tDimension reduction to p dimension to generate subspace S_sAnd S_t，G_d×pFor the set of all p-dimensional subspaces, each subspace is considered as the Grassman manifold space G_d×pAt one point above, let Φ (t) be G_d×pA geodesic line of (1), wherein t ∈ [0,1 ]]，Φ(0)＝S_sAnd Φ (1) is S_tAs ends of a geodesic line, z_iAnd z_jIs x_iAnd x_jThe inner product of the projected feature vector in the infinite dimensional space is expressed as:

in the above formula x_i,x_j∈D_s∪D_tAnd G denotes a geodesic core:

in the above formula R_sFrom X_sThe d-p dimensional feature composition, U, remaining after the P dimensional feature extraction by PCA₁And U₂Are orthogonal matrices and are solved by singular value decomposition₁,Λ₂And Λ₃For a diagonal matrix, the element values in the matrix are respectively:

mixing X_sAnd X_tSample feature space usage after projection

And

is shown in which

Step 23) Using Z_sTraining a classifier by a k nearest neighbor algorithm with k equal to 1, and then converting Z into Z_tBrought into the classifier to obtain pseudo labels

Step 24) Using Z_sAnd Z_tSelecting Gaussian kernels to construct kernel functions

Step 25) determining Z by k-nearest neighbor algorithm_sAnd Z_tThe neighbor relation of each point in the similarity matrix W is obtained:

in the above formula r (z)_i,z_j) 1 represents z_iAnd z_jAre in a mutual neighbor relationship; after obtaining W, calculating a Laplace matrix L:

in the above formula, D is a diagonal matrix consisting of

And calculating to obtain, wherein the manifold regular constraint term is expressed as:

step 26) for Z_sAnd Z_tPerforming conditional distribution self-adaptation, measuring the conditional distribution difference on a regeneration Hilbert space through the maximum mean difference, and approximating statistical estimation by using an empirical estimation formula:

where C represents the total number of classes of labels contained within the sample,

consists of two diagonal matrices, wherein:

represents Z_sAnd Z_tN with the middle label being l_lSet of classification results for individual samples, f denotes G_d×pThe following classification prediction function:

in the above formula, alpha ═ alpha₁,α₂,...,α_N]^TFor coefficient vectors, the final calculation formula of the conditional distribution adaptive term is:

MMD²(H_K,Q_s,Q_t)＝tr(f^TMf)

in the above formula, each element in M is directly calculated by the following formula:

according to the structure risk minimization principle, combining a manifold regularization term and a condition distribution self-adaptive term, and finally optimizing a target of a classifier f as follows:

in the above equation, U is an indication matrix of the domain where the sample is located, that is:

let the partial derivative of α in the optimization equation be 0 to solve for α:

α＝(λ₁I+(λ₂M+λ₃L+U)K)^-1UY^T

using f to complete pairs

Repeating step 26) e times to iteratively update M and α;

step 27) obtaining the prediction label of the sample of the batch

And storing the samples to wait for the input of the next batch of samples.

Images