A gas sensor drift compensation system and method based on multi-sensor information fusion

Abstract

The application provides a gas sensor drift compensation system and method based on multi-sensor information fusion. The system comprises a multi-sensor signal collector, a common feature extractor, a sub-domain alignment feature extractor, a sub-classifier and a voting strategy. The method designs a multi-sensor information fusion strategy at the decision level, and realizes drift compensation of the gas sensor by using the sub-domain adaptation idea, so as to improve the classification accuracy of the gas sensor array in the drift scene.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition technology, and in particular relates to a gas sensor drift compensation system and method based on multi-sensor information fusion. Background Technology

[0002] Gas sensors exhibit high sensitivity to flammable, toxic, and volatile organic compounds, and are currently widely used in environmental monitoring and food engineering. In practical applications, multiple gas sensors with different sensitivity characteristics are typically arranged into a gas sensor array, and pattern recognition algorithms are used to detect the gases.

[0003] Information fusion technology can extract useful information from multiple sensor data and perform information complementarity and optimized combination processing. Existing fusion technologies can be broadly classified into three categories: data-level fusion, feature-level fusion, and decision-level fusion. In the field of gas sensors, data-level fusion and feature-level fusion methods are commonly used to compress the original high-dimensional data and extract effective discriminative information, improving gas classification accuracy. Furthermore, gas sensors are susceptible to drift due to environmental temperature and humidity, which alters the correspondence between the gas sensor's input and output, thereby reducing the accuracy of gas classification results. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the present invention aims to provide a gas sensor drift compensation system and method based on multi-sensor information fusion. This method designs a multi-sensor information fusion strategy at the decision level and utilizes the subdomain adaptation concept to achieve drift compensation for the gas sensor, thereby improving the classification accuracy of the gas sensor array in drift scenarios.

[0005] The present invention is achieved through the following technical solution: The present invention proposes a gas sensor drift compensation system based on multi-sensor information fusion. The system includes a multi-sensor signal acquisition unit, a common feature extractor, a subdomain aligned feature extractor, a subclassifier, and a voting strategy.

[0006] Multi-sensor signal acquisition unit: N data points simultaneously acquired by N gas sensors represent the same state, i.e. The collected N samples, i.e., the N gas sensor signals, will be input into a common feature extractor for feature extraction.

[0007] Common Feature Extractor: The common feature extractor C(·) is used to extract common features from different sensors; it is a shared network structure consisting of three “one-dimensional convolutional layers + max pooling layers”.

[0008] Subdomain Alignment Feature Extractor: Each sensor signal corresponds to a subdomain alignment feature extractor D.i and the structure of these extractors are all the same; the sub-domain alignment feature extractor consists of three fully connected layers;

[0009] Sub-classifier: for each sensor signal, there is a sub-classifier CLS i and the structure of all the classifiers are the same; the sub-classifier CLS i consists of 3 fully connected layers, and its input is the output D of the sub-domain alignment feature extractor i (C(·)), the number of neurons in the output layer is equal to C, and the activation function is Softmax;

[0010] Voting strategy: according to the loss after training, the prediction results of each sub-classifier are assigned weights to give the final decision result of the target domain samples.

[0011] Further, the sub-domain alignment feature extractor is used to project the data from the source domain and the target domain to a specific feature space, specifically, for sensor i, the sensor signal from the source domain and the sensor signal from the target domain are processed by the common feature extractor, and the outputs are and Then, and are input into the sub-domain alignment feature extractor and projected to a specific feature space; the distribution difference between the source domain and the target domain in the specific feature space is estimated by formula (1), and the sub-domain feature distribution is aligned by minimizing the distribution difference during model training, thereby improving the gas classification accuracy in the drift scenario.

[0012]

[0013] where C represents the total number of gas categories, ||·||2 represents the 2-norm, and represent the set of source domain samples belonging to the c-th category and the set of target domain samples belonging to the c-th category in sensor i, respectively; and represent and the outputs of the common feature extractor and the sub-domain alignment feature extractor, respectively, and represent and the weights in the feature distribution loss For the source domain, where N Sc represents the number of samples belonging to class c in the source domain; for the target domain, the target domain samples have no real labels, in which case, it is assumed that represents the prediction result of the sub-classifier i on the sample , which is a probability distribution with a sum of 1; the distribution difference loss represents the average distance between the class centers of the source domain and the class centers of the target domain in the feature space, and if the distance is small enough, it is considered that the sub-domain alignment feature extractor has the function of aligning the feature distribution.

[0014] Further, for each sub-classifier, the classification loss of the source domain is:

[0015]

[0016] wherein, is equal to 0 or 1, if the real label of is class c, then Otherwise represents the probability of the sub-classifier predicting as a sample of class c.

[0017] Further, the difference between the classification results of the source domain and the target domain is:

[0018]

[0019]

[0020] Further, the calculation formula of the weight assigned to each sub-classifier is as follows:

[0021]

[0022] wherein

[0023]

[0024]

[0025] the target domain sample The final classification result is:

[0026]

[0027] wherein, represents the prediction result of the sub-classifier i on , that is, From the comprehensive formulas (5)-(8), it can be seen that the greater the loss of the network in which the sensor i is located, the smaller the influence of the sensor i on the final classification result, and the smaller the loss, the greater the contribution of the sensor i to the final classification result.

[0028] The application provides a compensation method of the gas sensor drift compensation system based on multi-sensor information fusion.

[0029] Step 1: collecting source domain data and target domain data, wherein the source domain is a clean multi-gas sensor response signal, and the target domain is a multi-gas sensor response signal under drift interference;

[0030] Step 2: inputting the labeled source domain data and the unlabeled target domain data into a network for training, optimizing network parameters by using and to establish a gas classification model with drift compensation capability;

[0031] Step 3: detecting the target domain data by using the trained gas classification model, and realizing gas classification according to a voting strategy of formula (8).

[0032] Further, in the method, the loss function is composed of four parts, including a prediction loss a distribution difference loss a source domain-based classifier difference loss and a target domain-based classifier difference loss The calculation formula is as follows:

[0033]

[0034] wherein, the prediction accuracy of each classifier can be reflected, the distribution difference between the source domain and the target domain can be quantitatively evaluated, and the degree of multi-sensor information fusion can be represented.

[0035] The application has the following beneficial effects:

[0036] The application provides a gas sensor drift compensation system and method based on multi-sensor information fusion. The method designs a multi-sensor information fusion strategy at the decision level, and realizes drift compensation of the gas sensor by using the sub-domain adaptation idea, so that the classification accuracy of the gas sensor array in the drift scene is improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a source domain data schematic diagram;

[0038] Figure 2 is a framework diagram of the gas sensor drift compensation system based on multi-sensor information fusion.

[0039] Figure 3The compensation method flow chart of the gas sensor drift compensation system based on multi-sensor information fusion. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0041] Suppose there are N different sensors, and the source domain D S is composed of N sensor data, that is, and and respectively represent the instance and label of the jth sample from the ith sensor. Since the source domain data D S is collected simultaneously by different sensors, the number of samples in each sensor data is the same. The unlabeled target domain is and where represents the instance of the jth sample from the ith sensor, and N T represents the number of samples of each . The objective of the present application is to reduce the influence of sensor drift and effectively fuse multi-sensor information, thereby improving the accuracy of gas detection in the drift scenario. The schematic diagram of the source domain data is shown in Figure 1 .

[0042] The present application takes gas sensors as the research object, and proposes a gas sensor drift compensation system based on multi-sensor information fusion. The system includes a multi-sensor signal collector, a public feature extractor, a sub-domain alignment feature extractor, a sub-classifier and a voting strategy; as shown in Figure 2 .

[0043] Multi-sensor signal collector: N data obtained by simultaneously collecting N gas sensors represent the same state, that is, The network structure of the present application is a deep learning framework that can automatically extract deep features from raw data, overcoming the serious dependence on prior knowledge of traditional methods; the N samples collected, i.e. N gas sensor signals, will be respectively input into the public feature extractor for feature extraction;

[0044] Common feature extractor: the common feature extractor is used to extract common features of different sensors; it is a shared network structure composed of three structures of "one-dimensional convolution layer + maximum pooling layer"; the former is suitable for processing sensor signals that can only move in one dimension, and the latter can remove redundant information and simplify network complexity;

[0045] Sub-domain alignment feature extractor: each sensor signal corresponds to a sub-domain alignment feature extractor D i , and the structures of these extractors are completely the same; the sub-domain alignment feature extractor is composed of three fully connected layers;

[0046] In the gas sensor drift scenario, even if the same sensor collects signals of the same state, their data distributions are different. This difference will reduce the classification accuracy of the gas classification model. In order to reduce the distribution difference, the sub-domain alignment feature extractor is used to project the data from the source domain and the target domain to a specific feature space. Specifically, for sensor i, the sensor signal from the source domain and the sensor signal (n≤N T ) are processed by the common feature extractor to output and respectively, and then and are input into the sub-domain alignment feature extractor and projected to a specific feature space; formula (1) is used to estimate the distribution difference between the source domain and the target domain in the specific feature space, and the sub-domain feature distribution is aligned by minimizing the distribution difference during model training, thereby improving the gas classification accuracy in the drift scenario;

[0047]

[0048] where C represents the total number of gas categories, ||·||2 represents the 2-norm, and represent the set of source domain samples belonging to the c-th category and the set of target domain samples belonging to the c-th category in sensor i, respectively; and represent the outputs of and after sequentially passing through the common feature extractor and the sub-domain alignment feature extractor, respectively, and represent the outputs of and in the feature distribution loss , and where denotes the number of samples in the source domain that belong to class c. denotes the prediction of sub-classifier i for sample , which is a probability distribution that sums up to 1. denotes the probability distribution of sub-classifier i predicting class c. In essence, the proposed distribution discrepancy loss denotes the average distance between the class centers of the source domain and the target domain in the feature space.

[0049] Sub-classifier: for each sensor signal, there is a sub-classifier CLS i , and all the classifiers have the same structure; the sub-classifier CLS i is composed of 3 fully connected layers, the input of which is the output D i (C(·)) of the sub-domain alignment feature extractor, the number of neurons of the output layer is equal to C, and the activation function is Softmax;

[0050] For each sub-classifier, the classification loss of the source domain is:

[0051]

[0052] wherein, is equal to 0 or 1, if the real label of is class c. denotes the probability of sub-classifier predicting class c for sample

[0053] The present application hopes that the prediction results of different sub-classifiers for the same time collected samples are as same as possible. Therefore, in the present application, the difference of the classification results of the source domain and the target domain is considered respectively as:

[0054]

[0055]

[0056] It can be seen that formula (3) and formula (4) are a multi-sensor signal fusion method, which realizes the effective fusion of different network structures at the decision level, and avoids the shortage that the traditional feature level fusion method seriously depends on prior knowledge.

[0057] Voting strategy: In order to improve the synergistic ability of multiple sub-classifiers, a voting strategy for multiple classifiers is proposed. It adaptively assigns weights to the prediction results of each sub-classifier according to the loss after training, thereby giving the final decision result of the target domain sample. It not only enhances the contribution of good classifiers, but also suppresses the effect of bad classifiers.

[0058] The calculation formula of the weight assigned to each sub-classifier is as follows:

[0059]

[0060] Among them

[0061]

[0062]

[0063] Then the target domain sample The final classification result is:

[0064]

[0065] Among them, The prediction result of sub-classifier i to , that is, From the comprehensive formula (5)-(8), it can be seen that the greater the loss of the network where sensor i is located, the smaller the influence of the final classification result, and the smaller the loss, the greater the contribution to the final classification result.

[0066] In the gas classification task under the drift scenario, the existing classification model usually obtains a low classification accuracy. On the other hand, designing a suitable multi-sensor fusion strategy is a challenge in the field of machine olfaction. In addition, it is essential to reasonably evaluate the classification results of multiple classification models based on different sensors and obtain reliable collaborative classification results. Therefore, the present application proposes a compensation method based on the gas sensor drift compensation system based on multi-sensor information fusion, which comprises:

[0067] Step 1: Collect source domain data and target domain data, wherein the source domain is clean multi-gas sensor response signal, and the target domain is multi-gas sensor response signal under drift interference;

[0068] Step 2: input the labeled source domain data and unlabeled target domain data into the network for training, and use and to optimize the network parameters, and establish a gas classification model with drift compensation ability;

[0069] Step 3: The target domain data is detected by using the trained gas classification model, and the gas classification is realized according to the voting strategy of formula (8).

[0070] In the method, the loss function is composed of four parts, including the prediction loss the distribution difference loss the source domain-based classifier difference loss and the target domain-based classifier difference loss The calculation formula is as follows:

[0071]

[0072] Among them, the prediction accuracy of each classifier can be reflected, the distribution difference between the source domain and the target domain can be quantitatively evaluated, and the degree of multi-sensor information fusion can be represented. Therefore, in the training stage, by minimizing the feature distribution of the source domain and the target domain is aligned, the prediction accuracy of each classifier is improved, and the prediction result difference between the classifiers is reduced. In the test stage, after the target domain sample is fed into the trained model, each sub-classifier will output a prediction result, then the trained loss is used to assign weights to these prediction results according to formula (7), and finally the final gas classification result is calculated according to the voting strategy of formula (8).

[0073] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.

Claims

1. A gas sensor drift compensation system based on multi-sensor information fusion, characterized in that: The system includes a multi-sensor signal acquisition unit, a common feature extractor, a subdomain aligned feature extractor, a subclassifier, and a voting strategy. Multi-sensor signal acquisition unit: N data points simultaneously acquired by N gas sensors represent the same state, i.e. The collected N samples, i.e., the N gas sensor signals, will be input into a common feature extractor for feature extraction. Common Feature Extractor: The common feature extractor C(·) is used to extract common features from different sensors; It is a shared network structure consisting of three "one-dimensional convolutional layers + max pooling layers"; Subdomain Alignment Feature Extractor: Each sensor signal corresponds to a subdomain alignment feature extractor D. i Furthermore, all of these extractors have identical structures; the subdomain aligned feature extractor consists of three fully connected layers. Subclassifier: For each sensor signal, there is a corresponding subclassifier CLS. i Furthermore, all classifiers have the same structure; the subclassifier CLS i It consists of three fully connected layers, and its input is the output D of the subdomain aligned feature extractor. i (C(·)), where the number of neurons in the output layer is equal to C, and the activation function is Softmax; Voting strategy: Based on the training loss, weights are adaptively assigned to the prediction results of each sub-classifier, thereby giving the final decision result for the target domain samples; The subdomain aligned feature extractor is used to project data from the source domain and the target domain onto a specific feature space. Specifically, for sensor i, the sensor signal from the source domain... and sensor signals from the target domain After processing by the common feature extractor, the outputs are as follows: and Then, and The input is projected into a subdomain aligned feature extractor and projected into a specific feature space; Equation (1) is used to estimate the distribution difference between the source domain and the target domain in the specific feature space, and the subdomain feature distribution is aligned by minimizing the distribution difference during model training, thereby improving the gas classification accuracy in drift scenarios; Where C represents the total number of gas categories, and ||·||2 represents the 2-norm. and Let represent the set of samples in sensor i that belong to the c-th class of the source domain and the c-th class of the target domain, respectively; and They represent and The outputs of the common feature extractor and the subdomain aligned feature extractor are processed sequentially. and They represent and In characteristic distribution loss The weights in the domain; for the source domain, in This represents the number of samples belonging to class c in the source domain; for the target domain, the samples have no true labels. In this case, it is assumed that... Subclassifier i represents the number of samples. The prediction result is a probability distribution with a sum of 1; the distribution difference loss is... This represents the average distance between the class centers of the source domain and the class centers of the target domain in the middle of the features. If this distance is small enough, the subdomain aligned feature extractor is considered to have the function of feature distribution alignment.

2. The system according to claim 1, characterized in that, For each sub-classifier, the classification loss of the source domain is: in, It equals 0 or 1, if If the actual label is class C, then otherwise Subclassifier prediction The probability of being a sample of class c.

3. The system according to claim 2, characterized in that, The differences between the classification results of the source domain and the target domain are as follows:

4. The system according to claim 3, characterized in that, The formula for calculating the weights assigned to each sub-classifier is as follows: in Then the target domain sample The final classification result is: in, Indicates subclassifier i pairs The prediction results, i.e. As can be seen from the combined equations (5)-(8), the greater the loss of the network where sensor i is located, the smaller its impact on the final classification result; the smaller the loss, the greater its contribution to the final classification result.

5. A compensation method for a gas sensor drift compensation system based on multi-sensor information fusion according to claim 4, characterized in that: The method includes: Step 1: Collect source domain data and target domain data, where the source domain is the clean multi-gas sensor response signal and the target domain is the multi-gas sensor response signal under drift interference; Step 2: Input labeled source domain data and unlabeled target domain data into the network for training, utilizing... and To optimize network parameters and establish a gas classification model with drift compensation capability; Step 3: Use the trained gas classification model to detect the target domain data and classify the gas according to the voting strategy of Equation (8).

6. The method according to claim 5, characterized in that, In the method, the loss function l total It consists of four parts, including loss prediction. Distributional difference loss Source domain-based classifier difference loss And target domain-based classifier difference loss The calculation formula is as follows: in, It can reflect the prediction accuracy of each subclassifier. It can quantitatively assess the distributional differences between the source and target domains. and It can represent the degree of information fusion from multiple sensors.

Images