Turbulent mixed gas recognition method based on signal reconstruction fine composite moving average fluctuation dispersion entropy

Abstract

The application discloses a turbulent mixed gas identification method based on signal reconstruction fine composite sliding average fluctuation dispersion entropy, and belongs to the gas detection field. The method sequentially executes the following steps: step a, reconstructing a time domain signal, first splicing the response signals of n sensors in a sensor array into a time sequence signal in sequence; step b, extracting features of the reconstructed time sequence signal by using a fine composite sliding average fluctuation dispersion entropy (RCMAFDE), and obtaining a turbulent mixed gas feature set; and step c, then constructing an MRVM classifier to identify the gas category, and selecting a sparrow optimization algorithm to optimize the kernel parameters of the Gaussian kernel function of the MRVM in order to improve the identification precision. The application realizes sufficient extraction of the gas category features by using the RCMAFDE as a feature extraction mode, improves the classification and identification precision, and the classification precision reaches 98.3%.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of gas detection and relates to a turbulent mixed gas identification method based on signal reconstruction fine composite moving average fluctuation dispersion entropy. BACKGROUND

[0002] In the coal mining industry, oil extraction and natural gas transmission pipes, a large number of flammable and explosive and toxic gases such as ethylene, methane, carbon monoxide and hydrogen are often accompanied. Real-time monitoring of the types and concentrations of these gases plays a crucial role in the safety and health of mining personnel and can effectively prevent accidents. In recent years, gas detection technology has developed rapidly and has potential application potential in the fields of aerospace, public safety, atmospheric environment detection, food quality monitoring and the like.

[0003] Current detection methods for mixed gases, such as spectral analysis, mass spectrometry, chromatography, etc., have high gas detection accuracy, but due to high cost, long response time and inability to measure online, etc., they cannot be popularized. Metal oxide semiconductor (MOS) gas sensors have unique advantages such as small size, low power consumption and high sensitivity, and have quickly occupied the gas sensor market since their appearance. Due to the cross-sensitivity problem of single gas sensor, it is easy to be disturbed by other unknown gases, so multiple MOS gas sensors are usually used to form a gas sensor array to solve the problem of mixed gas recognition and concentration estimation.

[0004] The classification research of mixed gas is applied to many fields, and the classification algorithms proposed are also various. Such as principal component analysis (PCA), support vector machine (SVM), artificial neural network (ANN), K nearest neighbor method (KNN), ensemble learning, etc. Mixed gas classification algorithms all show good mixed gas classification detection characteristics. Common mixed gas concentration prediction algorithms include support vector machine regression, BP neural network, linear regression, etc.

[0005] For gas qualitative identification, although the above methods can effectively realize the identification of mixed gas to some extent, PCA is a linear feature extraction method, and the extracted features are the linear sum of the original features. However, the response signal of the MOS gas sensor has nonlinear characteristics. Therefore, the linear feature extraction method such as PCA is not ideal. Artificial neural network (ANN) algorithm needs a large number of parameters, and neural network algorithm needs a large number of learning processes, which may not be feasible for small sample data sets. Ensemble learning needs a large amount of data and is not suitable for small sample data sets. SUMMARY

[0006] The technical problem solved by the present application is that the data collected by the MOS sensor array has a nonlinear phenomenon, and the existing feature extraction method does not extract sufficient features.

[0007] The technical solution adopted by the present application to solve its technical problem mainly includes the following steps:

[0008] Step a, reconstructing the time domain signal, first splicing the signals of the 8 sensors into a time sequence signal;

[0009] Step b, extracting features from the reconstructed time sequence signal using RCMAFDE to obtain a set of features of the turbulent mixed gas;

[0010] Step c, then using the MRVM classifier to identify the gas, in order to improve the identification accuracy, selecting the sparrow optimization algorithm to optimize the kernel parameters of the Gaussian kernel function of the MRVM;

[0011] The above-mentioned turbulent mixed gas identification method based on signal reconstruction and fine composite moving average fluctuation dispersion entropy, step a specifically includes the following steps:

[0012] Step a1, the response of the sensor array to the gas is saved in a.txt file, which is converted into a.mat file format matrix by matlab.

[0013] Step a2, the matrix size is m*n, where m represents the number of sampling points, and n represents the number of sensors, and the matrix is transposed to an n*m matrix.

[0014] Step a3, the n*m matrix is converted into a 1*M matrix (M=n*m), forming a new time domain signal.

[0015] Step a4, loop through N.txt files (N represents the number of gas samples), sequentially through steps a1, a2, a3, and finally form an N*M matrix.

[0016] The above-mentioned turbulent mixed gas identification method based on signal reconstruction and fine composite moving average fluctuation dispersion entropy, step b specifically includes the following steps:

[0017] Step b1, the N*M matrix obtained after the sample goes through step a, after RCMAFDE feature extraction, forms an N*Y matrix (Y represents the dimension of feature extraction).

[0018] The above-mentioned turbulent mixed gas identification method based on signal reconstruction and fine composite moving average fluctuation dispersion entropy, step c specifically includes the following steps:

[0019] Step c1, initialize the MRVM classifier, and select the Gaussian kernel function as the kernel function of the MRVM.

[0020] Step c2, select the sparrow optimization method to optimize the kernel parameters of the MRVM, and improve the recognition accuracy.

[0021] The beneficial effects of the present application are:

[0022] Through the data preprocessing operation, the influence of sensor drift is reduced, and the error of the experiment is reduced. By splicing the signals of the 8 sensors into a time sequence signal, the reconstructed time sequence signal is extracted by RCMAFDE, and the characteristics are fully extracted, solving the problem of nonlinear response of the sensor array to the gas. In addition, MRVM is selected as the classifier, and MRVM can directly output multiple types by using multiple probability likelihood functions. The sparrow optimization (SSA) method is used to optimize the kernel parameters of the MRVM, and the recognition accuracy is improved, and the accuracy reaches 98.335%. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a flow chart of a turbulent mixed gas recognition method based on signal reconstruction and wavelet dispersion entropy of fine composite moving average. Figure 1

[0024] Figure 2 is a time sequence signal graph obtained by splicing the signals collected by 8 sensors (taking a mixed gas of carbon monoxide with a concentration of 460 ppm and ethylene with a concentration of 96 ppm as an example). Figure 2

[0025] Figure 3 is a sliding average process of the sliding average factor τ=2 and τ=3. Figure 3

[0026] Figure 4 is a flow chart of RCMAFDE. Figure 4

[0027] Figure 5 is a flow chart of a mixed gas recognition method. Figure 5 DETAILED DESCRIPTION

[0028] The specific embodiment of the present application will be further described in detail below in combination with the drawings.

[0029] The turbulent mixed gas recognition method based on variational mode decomposition and amplitude perception permutation entropy in the specific embodiment comprises the following steps:

[0030] Step a, reconstructing the time domain signal, first splicing the signals of n sensors in the sensor array into a time sequence signal;

[0031] ​​​​​In this specific embodiment, a publicly available dataset is used, which is a publicly available dataset from Alexander Vergara et al. of the University of California, San Diego, California, USA, published in 2014. It contains three components: methane, ethylene, and carbon monoxide.

[0032] Dataset: An open sampling system was constructed for the experiment. The experimental setup consisted of a wind tunnel with two independent gas sources and a sensor array detection platform with eight built-in MOS gas sensors. Under the action of the fan, the gas continuously dispersed in the wind tunnel, and the concentration level changed continuously over time. To accurately estimate the gas concentration level at the sensor array, the experiment used micro-gas chromatography-superconductivity with ISQ single quadrupole mass spectrometry (Thermo Scientific) to estimate the gas concentration at the sensor array location. Different gas plumes corresponded to different gas concentration information. The mixed gas dataset was further refined into 30 mixed gas categories under the composition categories of ethylene, methane, and carbon monoxide, based on different gas concentration levels, as shown in Table 1.

[0033] Table 1 Dataset Category Information Table

[0034]

[0035] Step a1: The sensor array's response to the gas is saved in a .txt file. The .txt file is then converted into a matrix in .mat file format using MATLAB.

[0036] Step a2: The matrix size is m×n, where m represents the number of sampling points and n represents the number of sensors. Transpose the matrix to obtain an n×m matrix.

[0037] Step a3: Transform the n×m matrix into a 1×M matrix (M=n×m) to form a new time-domain signal.

[0038] Step a4: Iterate through N .txt files (N represents the number of gas samples), and then proceed through steps a1, a2, and a3 in sequence to finally form an N×M matrix.

[0039] Step b: Use RCMAFDE to extract features from the reconstructed time-series signal to obtain the feature set of turbulent mixed gas.

[0040] The specific calculation process of RCMAFDE is as follows:

[0041] (1). For a time series x = {x1, x2, ..., x...} N Construct the moving average sequence using the following formula.

[0042]

[0043] where τ = 1, 2, 3,... is the moving average factor. When τ = 1, the moving average sequence constructed is the original sequence.

[0044] When τ = 2 and τ = 3, their moving average processes are shown as Figure 3 It can be seen that in the RCMAFDE method, for each moving factor τ, τ different composite time series can be obtained.

[0045] (2) The moving average sequence of the original time series x = {x1, x2,..., x N} can be calculated as The probability average of the volatility dispersion pattern in the moving average sequence can be calculated as:

[0046]

[0047] In the formula, is the probability of the dispersion pattern in the moving average sequence .

[0048] When the moving average factor is τ, the RCMAFDE is defined as the Shannon entropy of the average dispersion model obtained by shifting the original time series:

[0049]

[0050] Fine composite moving average volatility dispersion entropy (RCMAFDE). The flow chart is shown in Figure 4

[0051] Step c, then use the MRVM classifier to identify the gas, in order to improve the identification accuracy, select the sparrow optimization algorithm to optimize the kernel parameters of the Gaussian kernel function of MRVM, step c specifically includes the following steps:

[0052] Step c1, initialize the MRVM classifier, and select the Gaussian kernel function as the kernel function of MRVM.

[0053] The specific implementation is:

[0054] Suppose the training sample set where x represents a D-dimensional input vector, t ∈ {1, 2, C} is the class label. In the regression problem, let the kernel function set be K ∈ R N×N , introduce an auxiliary variable y ∈ R C×N as the regression target, and introduce a weight parameter W ∈ R N×C , which can be obtained:

[0055] ​

[0056] where y cn is an element in y c is an element in w. N x (m,v) means x is normally distributed with mean m and variance v. Introducing multinomial probit link function, the goal of regression is converted into class label, as follows:

[0057]

[0058] Therefore, the multinomial probit likelihood function is obtained, as shown in the following formula, where u ~ N(0,1).

[0059]

[0060] In order to ensure the sparsity of the MRVM model, the weight parameter w is introduced with a standard normal prior distribution with mean 0 and variance , where nc is the prior parameter matrix A ∈ R N×C , and is subject to Gamma distribution with parameters τ and v. In order to ensure the sparsity of the model, τ and v are taken to be less than 10 -5 . The model of the multi-class relevance vector machine adopts a hierarchical Bayesian structure.

[0061] According to this structure, the posterior probability can be obtained.

[0062]

[0063] where A c is a diagonal matrix composed of c columns of A. Then, the maximum posterior likelihood estimation is obtained as follows:

[0064]

[0065] Step c2, as shown in the figure, the sparrow optimization method is selected to optimize the kernel parameter of MRVM, and the recognition accuracy is improved. Figure 5

[0066] In order to subjectively evaluate the classification effect, table 2 shows the recognition accuracy of the patent under the public data set.

[0067] Table 2 multi-classification accuracy of mixed gas of different training models

[0068] ​

Claims

1. A method for identifying a turbulent mixed gas based on signal reconstruction fine composite moving average fluctuation dispersion entropy, characterized in that, The method comprises the following steps: Step a, reconstructing the time domain signal, first splicing the response signals of n sensors in the sensor array into a time sequence signal; Step b, then using RCMAFDE to extract features from the reconstructed time sequence signal to obtain a turbulent mixed gas feature set; specifically The method comprises the following steps: The specific calculation process of RCMAFDE is as follows: Step bl, for a time series x = {xl, x2,..., x N}, construct a sliding average series y = {yl, y2,..., yT} with the equation Wherein τ = 1, 2, 3,... is a sliding average factor, when the value is τ = 1, the sliding average sequence constructed is the original sequence, in the RCMAFDE method, for each sliding factor τ, τ different composite time sequences are obtained; Step b2, a moving average sequence of the original time series x = {xl, x2,..., xN} Mid-wave discrete mode The probability average is calculated as: wherein is a moving average sequence discrete mode probability; Step b3, when the sliding average factor is τ, RCMAFDE is defined as the shannon entropy of the average discrete model obtained by shifting the original time sequence: Step b4, after the RCMAFDE feature extraction, the NXM matrix obtained after step a forms an NXY matrix, Y represents the dimension of feature extraction; Step c, then use MRVM classifier to identify gas, in order to improve the recognition accuracy, select sparrow optimization algorithm to optimize the kernel parameter of MRVM Gaussian kernel function.

2. The method of claim 1, wherein: In step a, the reconstructed time domain signal first splices the response signals of n sensors in the sensor array into a time sequence signal, specifically comprising the following steps: Step a1, the response of the sensor array to the gas is saved in a.txt file, which is converted into a.mat file format matrix by matlab; Step a2, the matrix size is m*n, where m represents the number of sampling points, n represents the number of sensors, and the matrix is transposed to an n*m matrix; Step a3, the n*m matrix is converted into a 1*M matrix, M = n*m, forming a new time domain signal; Step a4, loop through N.txt, N represents the number of gas samples, and sequentially pass through steps a1, a2 and a3 to finally form an N*M matrix.

3. The method of claim 1, wherein: In step c, MRVM classifier is used to identify gas, in order to improve the recognition accuracy, sparrow optimization algorithm is selected to optimize the kernel parameter of MRVM Gaussian kernel function, specifically comprising the following steps: Step c1, initialize MRVM classifier, select Gaussian kernel function as the kernel function of MRVM; Step c2, select sparrow optimization method to optimize the kernel parameter of MRVM, improve the recognition accuracy.

Images