Gas identification method based on sparse feature selection and spectrum enhancement network

Abstract

The invention provides a gas identification method based on sparse feature selection and a spectrum enhancement network, and relates to the technical field of deep learning and gas sensor intelligent detection, and the method comprises the steps: carrying out the data preprocessing through a gas sensor array data set, and obtaining the preprocessed gas sensor array data; sparse feature selection is carried out on the preprocessed gas sensor array data, and gas sensor array data of variable density sparse feature selection and a target label are obtained; constructing a spectrum enhancement network model; and according to the gas sensor array data selected by the variable density sparse features and the target label, utilizing the spectrum enhancement network model to obtain a gas identification result. According to the method, the cross sensitivity of the gas sensor facing the complex mixed gas is overcome, the performance of the deep learning model facing the complex mixed gas detection is improved, and the method has the characteristics of low calculation amount and high robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and intelligent detection of gas sensors, and particularly relates to a gas recognition method based on sparse feature selection and spectrum enhancement network. BACKGROUND

[0002] An artificial olfactory system is usually composed of two parts: a gas sensor array and a pattern recognition algorithm. Each sensor in the gas sensor array may produce cross-response to multiple gases, that is, the response signals of a single sensor to multiple different gases are similar. This cross-response phenomenon leads to a significant decrease in the selectivity and accuracy of traditional gas sensors in the detection of complex gas mixtures, which is specifically manifested in that the response of a sensor to a certain gas is not only affected by the gas itself, but also interfered by other gases, resulting in errors in the recognition result. This problem seriously limits the application effect of artificial olfactory systems in a multi-gas environment.

[0003] In order to overcome the cross-sensitivity of gas sensors and improve the performance of artificial olfactory systems, researchers have made a lot of efforts in the development and innovation of decision systems. In particular, with the continuous development of artificial intelligence technology, many studies have begun to explore how to use machine learning and deep learning methods to improve gas detection. However, these methods, although have made some progress in certain applications, still face some challenges, especially in the detection of complex mixed gases, how to effectively deal with the problem of sensor cross-response is still a key to be solved.

[0004] Traditional machine learning pattern recognition algorithms, such as support vector machine (SVM), k-nearest neighbor algorithm (k-NN), and random forest, although perform well in single gas recognition tasks, often fail to fully exert their advantages in the detection of complex gas mixtures. Since these algorithms usually rely on a large amount of prior knowledge and hand-designed feature engineering, they lack effective feature extraction and fusion capabilities when facing the cross-sensitivity of gas sensors, which limits their selectivity and accuracy in complex gas environments.

[0005] Deep learning algorithms, especially convolutional neural networks (CNN) and recurrent neural networks (RNN), have made significant achievements in gas detection due to their automatic feature extraction and powerful pattern recognition capabilities. However, these deep learning models mostly rely on single-modal data features, and lack effective feature engineering methods to handle the cross-response between sensors. Therefore, when facing complex mixed gases, they are easily disturbed by redundant features, making it difficult to effectively extract key features, and thus affecting the detection performance of the system.

[0006] Therefore, there is an urgent need for a gas recognition method based on sparse feature selection and spectrum enhancement network to solve the problems in the prior art. SUMMARY

[0007] The present application aims to provide a gas recognition method based on sparse feature selection and spectrum enhancement network, which aims to deeply mine the core features of gas sensor array data through feature engineering and network model collaborative optimization, overcome the cross-sensitivity of gas sensors when facing complex mixed gas, and improve the performance of deep learning models when facing complex mixed gas detection, with low computational complexity and strong robustness.

[0008] To achieve the above purpose, the present application provides a gas recognition method based on sparse feature selection and spectrum enhancement network, comprising the following steps:

[0009] S1, data preprocessing is performed on the gas sensor array data set to obtain preprocessed gas sensor array data;

[0010] S2, sparse feature selection is performed on the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels;

[0011] S3, a spectrum enhancement network model is constructed;

[0012] S4, the spectrum enhancement network model is used to obtain gas recognition results according to the variable density sparse feature selected gas sensor array data and the target labels.

[0013] Optionally, data preprocessing is performed on the gas sensor array data set to obtain preprocessed gas sensor array data, which comprises:

[0014] S1-1, the sensor array data of the mixed gas and the corresponding gas concentration and type labels are obtained as the gas sensor array data set;

[0015] S1-2, the preprocessed gas sensor array data is obtained by sequentially performing normalization and linear interpolation on the gas sensor array data set.

[0016] Optionally, sparse feature selection is performed on the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels, which comprises:

[0017] S2-1, the preprocessed gas sensor array data is divided based on the Monte Carlo cross-validation method to obtain the division data of the preprocessed gas sensor array data;

[0018] S2-2, obtaining gas instantaneous energy distribution characteristics of the gas sensor array data by using the divided data of the preprocessed gas sensor array data;

[0019] S2-3, performing sparse feature selection on the gas instantaneous energy distribution characteristics of the gas sensor array data to obtain variable-density sparse feature selected gas sensor array data and target labels.

[0020] Optionally, obtaining gas instantaneous energy distribution characteristics of the gas sensor array data by using the divided data of the preprocessed gas sensor array data comprises:

[0021] S2-2-1, performing complete set empirical mode decomposition on the divided data of the preprocessed gas sensor array data to obtain a set of gas intrinsic mode functions;

[0022] S2-2-2, performing Hilbert transform according to the set of gas intrinsic mode functions to obtain gas instantaneous energy distribution characteristics of the gas sensor array data.

[0023] Optionally, performing sparse feature selection on the gas instantaneous energy distribution characteristics of the gas sensor array data to obtain variable-density sparse feature selected gas sensor array data and target labels comprises:

[0024] S2-3-1, obtaining feature energy guided sparse selection probability distribution by using the gas instantaneous energy distribution characteristics of the gas sensor array data;

[0025] S2-3-2, constructing a variable-density sparse feature selection matrix according to the feature energy guided sparse selection probability distribution;

[0026] S2-3-3, performing variable-density sparse feature selection and enhancement on the divided data of the preprocessed gas sensor array data by using the variable-density sparse feature selection matrix to obtain variable-density sparse feature selected gas sensor array data;

[0027] S2-3-4, obtaining corresponding labels as target labels according to the variable-density sparse feature selected gas sensor array data.

[0028] Optionally, the spectrum enhancement network model is stacked by a spectrum enhancement multi-scale TCN feature extraction layer, a linear projection layer, N FFT Fourier encoding layers, and a feature decoding layer of a double-branch structure.

[0029] The spectrum enhancement multi-scale TCN feature extraction layer is composed of M TCN time series convolutional networks and a spectrum enhancement layer.

[0030] The Fourier encoding layer includes a fast Fourier transform layer, layer normalization, a residual connection layer, and a feedforward neural network.

[0031] The feature decoding layer of the double-branch structure includes a plurality of fully connected layers and Dropout layers, and is configured with a SiLU activation function.

[0032] Wherein, N FFT is a hyperparameter for controlling the number of Fourier encoder layers, M TCN is a hyperparameter for controlling the number of time convolution networks of different convolution kernel sizes, each of which includes a causal convolution layer and an expanded causal convolution layer, and is configured with a batch normalization layer and a ReLU activation function, and each of which is inserted with a plurality of spectral enhancement layers.

[0033] Optionally, constructing the spectral enhancement layer includes:

[0034] Performing adaptive average pooling operation on the one-dimensional feature sequence of the gas sensor array to obtain a feature sequence with a uniform sequence length;

[0035] According to the frequency index set corresponding to the uniform sequence length, a frequency index set is obtained.

[0036] Performing one-dimensional discrete cosine transform on the feature sequence with a uniform sequence length to obtain a frequency domain feature value;

[0037] Using the learnable weight parameters in the spectral domain to perform adaptive weighting on the frequency domain feature value based on the frequency index set to obtain an adaptively weighted and fused spectral feature;

[0038] According to the feature sequence with a uniform sequence length and the adaptively weighted and fused spectral feature, a spectral enhancement feature representation is obtained by using a fully connected layer and an activation function for feature enhancement, and then a spectral enhancement layer is constructed.

[0039] Wherein, the one-dimensional feature sequence of the gas sensor array is a one-dimensional feature sequence of the gas sensor array extracted by the causal convolution layer or the expanded causal convolution layer of the time convolution network.

[0040] Optionally, using the spectral enhancement network model to obtain a gas recognition result according to the gas sensor array data selected by the variable density sparse feature and the target label includes:

[0041] S4-1, using the gas sensor array data selected by the variable density sparse feature and the target label to train the spectral enhancement network model to obtain a trained spectral enhancement network model;

[0042] S4-2, inputting the variable-density sparse feature selection gas sensor array data into the trained spectrum enhancement network model, obtaining predicted features for characterizing gas species and concentration as a gas recognition result.

[0043] Compared with the closest prior art, the present application has the beneficial effects of:

[0044] The present application first extracts and enhances the core information of the data based on the instantaneous energy distribution characteristics of the gas sensor array, and then inputs the extracted core features into the constructed spectrum enhancement multi-scale Fourier network for gas species recognition and concentration prediction. Compared with traditional methods, the present application has the following significant advantages in data processing, feature extraction and network model design:

[0045] (1) The feature engineering method used in the present application can effectively screen and extract the most recognizable core features in the gas sensor array data, greatly reducing the redundancy of the data. Specifically, through the variable-density sparse feature selection mechanism guided by feature energy, only about 1 / 100 of the original data is used to identify and predict the performance, which is better than the end-to-end method of directly inputting the original data. This strategy not only effectively removes irrelevant features and improves pattern recognition effect, but also reduces the consumption of computing resources, significantly improving the response speed and processing efficiency of the system in practical application scenarios.

[0046] (2) The spectrum enhancement network proposed in the present application can effectively mine and dynamically enhance the spectral domain feature information of the gas sensor array data. Specifically, first, a spectrum enhancement multi-scale time series convolution network (TCN) is constructed to capture the features of the gas data at multiple time scales. Then, a dynamic spectrum enhancement mechanism based on instantaneous energy distribution is introduced on the feature sequence of each scale, that is, the discrete cosine transform (DCT) and the dynamic frequency component selection method are used to adaptively select key frequency components in the frequency domain and apply learnable frequency weights to them to highlight the core spectral features in the data. Further, the Fourier encoder realizes the deep coding and fusion of the features in the spectral domain, enhancing the network's ability to represent feature information. Finally, the decoding module precisely maps the high-dimensional features after spectrum enhancement to the output space of the gas species classification task and the gas concentration regression task, significantly improving the network's recognition performance and concentration prediction accuracy for complex mixed gases. Compared with traditional single-scale or shallow feature extraction methods, the spectrum enhancement network proposed in the present application not only has stronger multi-scale feature expression ability, but also can dynamically and data-drivenly enhance the key features in the spectral domain, better adapting to the high dynamics and cross-sensitivity problems in complex mixed gas environments, and overall improving the recognition accuracy and prediction robustness of the system.

[0047] (3) The application deeply cooperates and optimizes feature engineering and network model design, effectively realizes the organic combination of core feature extraction and model training. The powerful core feature extraction capability of the feature engineering link ensures that the core information with the most representative and discriminability in the data is effectively retained; and the spectrum enhancement network model further effectively solves the feature redundancy and interference problems caused by the cross-sensitivity of the gas sensor through the dynamic frequency component selection mechanism and the Fourier coding mechanism. This cooperative optimization strategy significantly improves the adaptability, stability and robustness of the system in complex gas environment, and shows obvious characteristics superior to traditional methods, especially in the application environment where the gas composition is complex and dynamically changes frequently, which shows higher generalization ability and prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings needed in the description of the specific embodiments or prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0049] Figure 1 A flow chart of a gas recognition method based on sparse feature selection and spectrum enhancement network according to an embodiment of the present application;

[0050] Figure 2 A schematic diagram of a gas instantaneous energy distribution feature extraction method according to an embodiment of the present application;

[0051] Figure 3 A sparse feature selection schematic diagram based on gas feature energy guidance according to an embodiment of the present application;

[0052] Figure 4 A spectrum enhancement network model structure diagram according to an embodiment of the present application;

[0053] Figure 5 A spectrum enhancement multi-scale TCN feature extraction layer structure diagram according to an embodiment of the present application;

[0054] Figure 6 A spectrum enhancement layer structure diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described below in a clear and complete manner with reference to the specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0056] The terms used in the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0057] Artificial olfactory systems have wide application value and important social significance in modern society, and their application fields cover environmental monitoring, industrial safety, food quality control, medical diagnosis and other aspects. For example, in environmental monitoring, artificial olfactory systems can detect the concentration of harmful gases in the air in real time and provide early warning information; in industrial production, the system can be used for gas leakage detection to ensure production safety; in the food industry, the system can be used to detect the freshness and preservation effect of food; in the medical field, by analyzing the volatile organic compounds (VOCs) in the patient's breath, artificial olfactory systems can assist in early diagnosis of diseases. These applications not only improve the efficiency and safety of various industries, but also play an important role in public health and environmental protection.

[0058] Artificial olfactory systems are usually composed of two parts: gas sensor array and pattern recognition algorithm, namely the acquisition system and the decision system. The gas sensor array is composed of multiple sensors, each of which can respond to different types of gases. The pattern recognition algorithm is responsible for processing and analyzing the data collected by the sensor array, and then realizing the accurate identification of gas types and concentrations. Currently, common gas sensors include metal oxide semiconductor (MOS), electrochemical sensors, optical sensors, etc., which have different characteristics and application ranges. The pattern recognition algorithm includes traditional machine learning algorithms (such as support vector machine, decision tree) and modern deep learning algorithms (such as convolutional neural network, recurrent neural network).

[0059] The sensors in a gas sensor array can produce cross responses to multiple gases, which leads to a significant decrease in the selectivity and accuracy of traditional gas sensors in detecting complex gas mixtures, severely limiting the application effect of artificial olfactory systems in multi-gas environments. In order to overcome the cross sensitivity of gas sensors and improve the performance of artificial olfactory systems, many studies have begun to explore how to use machine learning and deep learning methods to improve gas detection. However, these methods have made some progress in certain applications, but still face some challenges, especially in the detection of complex mixed gases, how to effectively deal with the problem of sensor cross response is still a key issue to be solved.

[0060] Although traditional machine learning pattern recognition algorithms perform well in single gas identification tasks, they often fail to fully exploit their advantages in the detection of complex gas mixtures. Since these algorithms usually rely on a large amount of prior knowledge and hand-designed feature engineering, they lack effective feature extraction and fusion capabilities when faced with gas sensor cross sensitivity, which limits their selectivity and accuracy in complex gas environments. Deep learning algorithms, especially convolutional neural networks (CNN) and recurrent neural networks (RNN), have made significant achievements in gas detection by virtue of their automatic feature extraction and powerful pattern recognition capabilities, but these deep learning models mostly rely on single modal data features and lack effective feature engineering methods to deal with the cross response between sensors. Therefore, when faced with complex mixed gases, they are easily disturbed by redundant features, making it difficult to effectively extract key features, which in turn affects the detection performance of the system.

[0061] Therefore, the key to improving the performance of artificial olfactory systems is the collaborative innovation of feature engineering and network models, i.e., designing effective feature engineering methods to extract the most core features of gas sensor array data, reducing data redundancy, removing features irrelevant to gas identification, and extracting core features with high recognition degree; then combining advanced deep learning models to further mine and learn complex patterns and potential relationships in core features, and collaboratively optimize pattern recognition algorithms to cope with the challenges brought by gas sensor cross response, and achieve high-accuracy and high-robustness gas type and concentration detection. In view of the above problems, the present application proposes a gas identification method based on sparse feature selection and spectral enhancement network, which extracts the core features of gas sensors based on sparse feature selection, and uses a deep learning model with good performance to overcome the cross sensitivity of gas sensors in mixed gas detection and improve the performance of artificial olfactory systems.

[0062] As shown in Figure 1 , the embodiment of the present application provides a gas identification method based on sparse feature selection and spectral enhancement network, comprising the following steps:

[0063] S1. Preprocess the gas sensor array dataset to obtain the preprocessed gas sensor array data.

[0064] S2. Perform sparse feature selection on the preprocessed gas sensor array data to obtain gas sensor array data with variable density sparse feature selection and target labels.

[0065] S3. Construct a spectrum enhancement network model;

[0066] S4. Based on the gas sensor array data selected according to the variable density sparse features and the target label, the gas identification result is obtained using the spectrum enhancement network model.

[0067] S1 specifically includes:

[0068] S1-1. Obtain the sensor array data of the mixed gas and the corresponding gas concentration and type labels as the gas sensor array dataset;

[0069] S1-2. Normalization and linear interpolation are performed sequentially on the gas sensor array dataset to obtain the preprocessed gas sensor array data.

[0070] This embodiment acquires sensor array data of mixed gases, along with corresponding gas concentration and type labels, as a gas sensor array dataset. The gas sensor array dataset is then preprocessed, specifically:

[0071] The dataset used in this embodiment is the Gas Sensor Array Underdynamic Gas Mixtures (Dataset 2). The sensor array in this dataset contains 16 chemical sensors of 4 types, with each type including 4 sensors: TGS-2600, TGS-2602, TGS-2610, and TGS-2620.

[0072] The data acquisition process is as follows: the sensor array is placed in a gas measurement chamber with a volume of 60 ml, the gas mixture to be tested is injected at a constant flow rate of 300 ml / min, and the electrical signal response data of the sensor array is continuously recorded at a sampling frequency of 100 Hz.

[0073] The dataset includes ethylene and carbon monoxide, both of which exhibited random concentration variations during the measurements. Ethylene concentrations ranged from 0 to 20 ppm, and carbon monoxide concentrations ranged from 0 to 600 ppm. Each measurement was performed by continuously recording data sequences from 16 sensor arrays, with each sequence recording for 12 hours. The gas was injected at a constant flow rate of 300 mL / min throughout the data sampling process. The sensor arrays acquired gas response signals in real-time at a sampling frequency of 100 Hz. During the experiment, all possible concentration variation patterns (including increases, decreases, and zero concentrations of a single gas, as well as constant and varying concentrations in mixed gas conditions) were comprehensively covered, resulting in diverse and complex multivariate sensor response data.

[0074] Because the sensor array used is highly sensitive to ethylene and carbon monoxide gases at low concentrations, and the response amplitudes of the sensors to different gases vary little, the task of identifying and predicting the concentration of mixed gases from gas sensor datasets presents a significant challenge. This embodiment, based on the gas sensor dataset, removes obviously faulty or severely drifting data segments, retaining only the desorption signals from the sensor array data for subsequent training and evaluation of the network model.

[0075] After the data is prepared, each data point is preprocessed, namely, normalization and linear interpolation. The calculation formulas for normalization and linear interpolation are as follows:

[0076]

[0077] Among them, D interp (p,q) represents the preprocessed gas sensor array data of the q-th gas sensor at time index p, D(p,q) represents the raw response value of the q-th gas sensor at time index p, D(u,q) represents the raw response value of the q-th gas sensor at time index u, L is the original sequence length of the single gas sensor array response data after unfolding along the time index dimension before normalization and interpolation, S is the total number of gas sensors, and Interp(.) represents the interpolation operation, specifically cubic spline interpolation, to improve the temporal resolution of the gas sensor array data and increase the speed of subsequent operations.

[0078] In this embodiment, each data set contains readings from 16 gas sensors with varying numbers of sampling points. After preprocessing, the number of sampling points for all data sets is interpolated to 20,000.

[0079] S2 specifically includes:

[0080] S2-1. Based on the preprocessed gas sensor array data, perform partitioning processing using the Monte Carlo cross-validation method to obtain partitioning data of the preprocessed gas sensor array data.

[0081] S2-2. Using the partitioned data of the preprocessed gas sensor array data, obtain the instantaneous energy distribution characteristics of the gas in the gas sensor array data;

[0082] S2-3. Perform sparse feature selection on the instantaneous energy distribution characteristics of the gas in the gas sensor array data to obtain gas sensor array data with variable density sparse feature selection and target label.

[0083] S2-2 specifically includes:

[0084] S2-2-1. Perform complete set empirical mode decomposition on the preprocessed gas sensor array data to obtain the set of gas intrinsic mode functions.

[0085] S2-2-2. Perform Hilbert transform based on the set of gas intrinsic mode functions to obtain the instantaneous energy distribution characteristics of the gas sensor array data.

[0086] S2-3 specifically includes:

[0087] S2-3-1. Utilize the instantaneous energy distribution characteristics of the gas in the gas sensor array data to obtain the sparse selection probability distribution guided by the characteristic energy.

[0088] S2-3-2. Construct a variable-density sparse feature selection matrix based on the sparse selection probability distribution guided by the feature energy.

[0089] S2-3-3. The variable density sparse feature selection matrix is ​​used to perform variable density sparse feature selection and enhancement on the preprocessed gas sensor array data to obtain gas sensor array data with variable density sparse feature selection.

[0090] S2-3-4. Based on the gas sensor array data selected according to the variable density sparse characteristics, obtain the corresponding label as the target label.

[0091] This embodiment extracts the instantaneous energy distribution features of the gas from the preprocessed gas sensor array data, constructs a variable density sparse feature selection matrix, and performs sparse feature selection and enhancement on the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data and corresponding labels. Specifically, the variable density sparse feature selected gas sensor array data is obtained by selecting features from the preprocessed data based on the sparse feature selection matrix to extract features from data points with different density distributions.

[0092] Monte Carlo cross-validation is used to partition the preprocessed gas sensor array data. For each channel of the partitioned data, adaptive noise-based complete ensemble empirical mode decomposition (CEEMDAN) and Hilbert transform are performed to extract the instantaneous gas energy distribution features. Based on these instantaneous gas energy distribution features, a feature energy-guided sparse selection probability distribution is constructed. A variable-density sparse feature selection matrix is ​​then constructed based on this probability distribution. Finally, the variable-density sparse feature selection matrix is ​​used to perform sparse feature selection and enhancement on the preprocessed gas sensor array partitioned data.

[0093] To ensure the comprehensiveness and generalization ability of the model performance evaluation, this embodiment uses the Monte Carlo cross-validation method to partition the preprocessed gas sensor array data, obtaining the partitioned data D of the preprocessed gas sensor array data. divide The mathematical expression for the data generation process is as follows:

[0094]

[0095] MonteCarloSplit(.) represents the operation of randomly partitioning the data based on the Monte Carlo cross-validation method. D represents the preprocessed subset of gas sensor array data generated by the i-th Monte Carlo cross-validation partition. intrep This represents the preprocessed data from the gas sensor array;

[0096] The specific process of the Monte Carlo cross-validation method is as follows:

[0097] A1. The preprocessed gas sensor array data is randomly divided into 10 non-overlapping subsets;

[0098] A2. For each iteration, randomly select 6 subsets as the training set and the remaining 4 subsets as the validation set;

[0099] A3. Save the training set generated after each partition. and verification set As the data for the i-th group;

[0100] A4. Repeat the above random partitioning and saving process 10 times to finally obtain 10 sets of training and validation data partitions, which will be used for subsequent sparse feature selection and model training.

[0101] like Figure 2 As shown, in this embodiment, the preprocessed gas sensor array data is divided into data D. divide Performing complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and Hilbert transform to extract the instantaneous gas energy distribution characteristics E from the gas sensor array data. norm (k), the specific steps are as follows:

[0102] B1. Division of preprocessed gas sensor array data D divide The CEEMDAN decomposition is performed using the following formula:

[0103]

[0104] Among them, D ceemdan (k) represents the set of gas intrinsic mode functions (IMFs) obtained by full set empirical mode decomposition of the preprocessed gas sensor array data, CEEMDAN(.) represents the full set empirical mode decomposition, and A represents the total number of IMFs obtained by CEEMDAN decomposition. a (k) represents the specific value of the a-th IMF at time index k.

[0105] In this embodiment, the specific value of A is 10.

[0106] B2. Obtaining the instantaneous energy distribution characteristics of gas from gas sensor array data based on the Hilbert transform using the set of gas intrinsic mode functions. norm (k), the calculation formula is as follows:

[0107]

[0108] Where S represents the total number of gas sensors, B r H(.) represents the total number of IMFs obtained after the r-th gas sensor is decomposed by CEEMDAN, and H(.) represents the Hilbert transform. b,r (k) represents the value of the b-th IMF corresponding to the r-th gas sensor at time index k, where IMF b,r(z) represents the value of the intrinsic mode function of the b-th gas corresponding to the r-th gas sensor at the auxiliary time index z, where z is the auxiliary time index used to find the global maximum value of the instantaneous energy of the gas, and ε is a small constant to prevent the denominator from being zero.

[0109] In this embodiment, a feature energy-guided sparse selection probability distribution is constructed based on the instantaneous energy distribution characteristics of the gas sensor array data. Variable density sparse feature selection and enhancement are then applied to the preprocessed gas sensor array data. Figure 3 As shown, the specific steps are as follows:

[0110] C1. Instantaneous energy distribution characteristics of gas based on gas sensor array data E norm (k), construct the sparse selection probability distribution P(k) guided by feature energy, and calculate it as follows:

[0111]

[0112] Where α is the nonlinear amplification hyperparameter used to enhance the selection probability of high-energy regions, x represents the auxiliary variable in the normalized denominator used to traverse the entire time series index, and E norm (x) represents the instantaneous energy distribution characteristic value of the gas at point x, obtained by normalization after the gas sensor array data is processed by CEEMDAN decomposition and Hilbert transform. It is used to reflect the local energy intensity of the gas response data at that time point.

[0113] In this implementation, α is set to 2.

[0114] C2. Based on the sparse selection probability distribution P(k) guided by feature energy, from the time index set {1,2,…,L} of the data sequence... s Sampling N in} s Using 1 independent index point, construct a variable-density sparse feature selection matrix M. u The calculation formula is as follows:

[0115] M u ={k1,k2,…,k N}k i ~P(k)i=1,2,…,N s

[0116] Where, k i M represents the i-th time index point randomly selected based on the sparse selection probability distribution P(k) guided by feature energy. u Contains N s A separate time index.

[0117] To minimize data redundancy and facilitate feature extraction in subsequent network models, this embodiment expects N sparse feature selection points.s Set to 16, the same as the number of sensors.

[0118] In this embodiment, for training data, a large batch of variable-density sparse feature selection matrices are generated based on the feature energy-guided sparse selection probability distribution for sparse feature selection and enhancement of the training data; for validation data, a single variable-density sparse feature selection matrix is ​​generated based on the feature energy-guided sparse selection probability distribution for feature selection.

[0119] For the training data, the process of constructing the variable density sparse feature selection matrix described in step C2 is repeated multiple times based on the feature energy-guided sparse selection probability distribution P(k), each time constructing an independent variable density sparse feature selection matrix M. u This approach generates a large number of different variable-density sparse feature selection matrices for the same original data. The selected feature index position is dynamically determined based on the distribution of data feature energy each time, ensuring that the enhanced data fully represents the core features of the original data. Randomness is also introduced to improve data diversity and generalization performance during model training. In this embodiment, step C216384 is repeated for the training data, generating 16384 variable-density sparse feature selection matrices.

[0120] C3. Using a variable density sparse feature selection matrix, perform sparse feature selection and enhancement on the preprocessed gas sensor array data to obtain gas sensor array data with variable density sparse feature selection. Feauture (f,g), the calculation formula is as follows:

[0121] D Feauture (f,g)=D divide (k f ,g)k f ∈M u f = 1, 2, ..., N s g = 1, 2, ..., S

[0122] Where f represents the new data time index after sparse selection, g represents the channel index of the gas sensor, and k f Choose matrix M for variable density sparse features u The f-th time index point in the sequence corresponds to the position D in the original data sequence. divide (k f (g) indicates that in the original partitioned data D divide In the middle, located at time index k f And the original preprocessed feature value at channel index g.

[0123] C4. Label the gas sensor array data with variable density sparse features. The label formats used in this embodiment for classification and regression tasks are shown in Table 1.

[0124] Table 1

[0125]

[0126] This embodiment employs an extremely unbalanced training and validation set ratio, using only 40% of the data as the training set. It verifies the ability of feature energy-guided variable density sparse feature selection to extract and enhance the core features of gas sensor array data, as well as the feature extraction and fusion capabilities of the spectrum enhancement network, in the case of feature loss.

[0127] Furthermore, the spectrum enhancement network model consists of a spectrum enhancement multi-scale TCN feature extraction layer, a linear projection layer, and N... FFT It consists of a stack of Fourier coding layers and a feature decoding layer with a dual-branch structure;

[0128] The spectrum-enhanced multi-scale TCN feature extraction layer consists of M TCN It consists of a temporal convolutional network and a spectral enhancement layer;

[0129] The Fourier coding layer includes a fast Fourier transform layer, layer normalization, residual connection layer, and feedforward neural network;

[0130] The feature decoding layer of the dual-branch structure includes several fully connected layers and Dropout layers, and is configured with the SiLU activation function;

[0131] Where, N FFT To control the hyperparameters of the number of layers in a Fourier encoder, M TCN To control the number of temporal convolutional networks with different kernel sizes, each temporal convolutional network includes a causal convolutional layer and a dilated causal convolutional layer, and is equipped with a batch normalization layer and a ReLU activation function. Furthermore, each temporal convolutional network is interpolated with several spectral enhancement layers.

[0132] Furthermore, constructing the spectrum enhancement layer includes:

[0133] An adaptive average pooling operation is performed using the one-dimensional feature sequence of the gas sensor array to obtain a feature sequence of uniform length.

[0134] Obtain the frequency index set based on the frequency index set corresponding to the uniform sequence length;

[0135] A one-dimensional discrete cosine transform is performed on the feature sequence with the unified sequence length to obtain frequency domain feature values.

[0136] By using learnable weight parameters in the spectral domain to adaptively weight the frequency domain feature values ​​based on the frequency index set, adaptive weighted fused spectral features are obtained.

[0137] Based on the feature sequence of the unified sequence length and the spectral features of the adaptive weighted fusion, feature enhancement is performed using a fully connected layer and activation function to obtain a spectral enhancement feature representation, and then a spectral enhancement layer is constructed.

[0138] Among them, the one-dimensional feature sequence of the gas sensor array is the one-dimensional feature sequence of the gas sensor array extracted by the causal convolutional layer or dilated causal convolutional layer of the temporal convolutional network.

[0139] This embodiment constructs a spectrum enhancement network model, specifically:

[0140] First, a spectrally enhanced multi-scale TCN feature extraction layer is constructed based on temporal convolutional networks (TCNs) with different kernel sizes and a spectral enhancement layer to extract feature information from gas sensor array data at different scales, and then multi-scale fusion is performed. Second, feature encoding and reconstruction are performed through a Fourier encoder. Finally, the feature information is mapped to the output space of the gas type classification task and the gas concentration regression task respectively for feature decoding.

[0141] like Figure 4 As shown, the spectrum enhancement network model used in this embodiment is used to enhance and encode / decode the input features in the spectrum domain. From top to bottom, it consists of a spectrum enhancement multi-scale TCN feature extraction layer, a linear projection layer, and an N... FFT The system consists of stacked Fourier coding layers and dual-branch feature decoding layers. The dual-branch feature decoding layer includes a classification branch and a regression branch. FFT To control the hyperparameter of the number of Fourier encoder layers, in this embodiment, N FFT Set to 4. The spectrum enhancement network model specifically includes the following:

[0142] D1, Spectrum Enhancement Multi-Scale TCN Feature Extraction Layer: (e.g.) Figure 5 As shown, M is constructed based on different convolution kernel sizes. TCN A parallel temporal convolutional network (TCN), where M TCN To control the number of TCNs with different kernel sizes, a hyperparameter is used. Each TCN contains two causal convolutional layers, two dilated causal convolutional layers, and is equipped with a batch normalization (BN) layer and a ReLU activation function for feature extraction. Each TCN is interpolated with several spectral enhancement layers to effectively utilize the spectral features of the data to enhance the information. Finally, the output features of TCNs with different kernel sizes are fused at multiple scales along the channel dimension to achieve time-frequency feature fusion at different scales.

[0143] D2, Linear Projection Layer: The fused features output from the spectral enhancement multi-scale TCN feature extraction layer are input into the linear projection layer. Based on the fully connected layer, the feature dimension is unified to meet the feature input requirements of the subsequent Fourier coding layer, thereby improving the network's expressive power and feature stability.

[0144] D3. Fourier Coding Layer: In each Fourier coding layer, the Fast Fourier Transform (FFT) is used to replace the multi-head attention mechanism in the traditional Transformer, realizing the conversion of gas data from the time domain to the frequency domain. Specifically, this includes:

[0145] First, the sequence data input to the FFT layer is subjected to FFT transformation to complete the frequency connection of gas features, thereby capturing the global correlation information of gas features in the frequency domain.

[0146] Secondly, by using layer normalization (abbreviated as Normalize in the figure) and introducing residual connection layers (abbreviated as add in the figure), the stability of the training process is ensured. Then, a feedforward neural network is used to further extract relevant features of the gas sensor data, thereby capturing the dynamic pattern of gas data changing over time.

[0147] Finally, layer normalization and residual connections are performed again to enhance the expressive power of features, achieve efficient fusion of gas dynamic temporal features and frequency domain features, improve the model's ability to identify gas types and concentrations, and enhance the accuracy of the network in gas type identification and concentration prediction.

[0148] D4. Feature Decoding Layer with Dual Branch Structure: A feature decoding structure is constructed using K fully connected layers (abbreviated as Full-connect in the diagram), a Dropout layer, and a SiLU activation function. This structure maps the high-dimensional features after Fourier encoding to the output spaces of the gas type classification task and the gas concentration regression task, respectively, achieving effective decoding and accurate prediction of the input data. Here, K is a hyperparameter controlling the number of fully connected layers; in this embodiment, K is set to 2.

[0149] like Figure 6 As shown, the construction of the spectrum enhancement layer specifically includes the following steps:

[0150] E1, the one-dimensional feature sequence X of the gas sensor array extracted through the TCN intermediate layer. tcn Adaptive average pooling is performed to unify the sequence length, ensuring consistency in the network input dimension, and thus obtaining a feature sequence X of uniform length. pool (c,t):

[0151] X pool(c,t)=AdaptivePool(X tcn (c,d)),t=1,2,…,T

[0152] Among them, X tcn (c,d) represents the one-dimensional feature sequence of the gas sensor array at the original time index d for the c-th sensor channel output by the TCN intermediate layer. AdaptivePool(.) represents the adaptive average pooling operation, where c is the channel index of the gas sensor array and t is the unified time index after adaptive average pooling.

[0153] E2. In this embodiment, considering that the TCN network structure has a strong response to high-frequency components during feature modeling, but is somewhat lacking in capturing low-frequency slow-changing trends, in order to improve the model's ability to perceive slowly changing gas signals, the preset frequency selection strategy is set to prioritize the selection of low-frequency components.

[0154] Specifically, based on the frequency index set {1,2,…,T} corresponding to the unified sequence length T, the first K frequency indices are selected sequentially in ascending order of frequency to form the frequency index set F. sel The calculation formula is as follows:

[0155] F sel ={f1,f2,…,f K},f i ∈{1,2,…,T},i=1,2,…,K

[0156] Among them, f i This indicates the i-th frequency index selected, and K represents the total number of frequency components ultimately selected.

[0157] E3, for characteristic sequences X of uniform sequence length pool (c,t) is subjected to a one-dimensional discrete cosine transform (DCT) to obtain the frequency domain eigenvalues ​​X. dct The calculation formula is as follows:

[0158] X dct (c,f)=DCT(X pool (c,t)),f=1,2,…,T

[0159] Among them, X dct (c,f) represents the frequency domain feature value of the c-th channel at the frequency index f after DCT transformation. DCT(.) represents one-dimensional discrete cosine transform.

[0160] E4. Utilizing learnable weight parameters w in the frequency domain f The selected frequency components are adaptively weighted to obtain the adaptively weighted fused spectral features X.freq The calculation formula is as follows:

[0161]

[0162] Among them, X freq (c) represents the spectral characteristics of the adaptive weighted fusion of the c-th channel. In this embodiment, the first four low-frequency components are selected for adaptive weighting.

[0163] E5, Spectral Features X Based on Adaptive Weighted Fusion freq Dynamic weights along the channel dimension are generated through a fully connected layer (FC) and a sigmoid activation function to dynamically enhance feature information. The generated dynamic weights are then multiplied element-wise with the original input feature sequence to obtain the feature representation after dynamic frequency component selection and adaptive enhancement, i.e., the feature-enhanced feature representation. The calculation formula is as follows:

[0164] X enhanced (c,t)=X pool (c,t)·Sigmoid(FC(X freq (c)))

[0165] Among them, X enhanced (c,t) represents the spectral enhancement feature, and Sigmoid(.) represents the activation function.

[0166] S4 specifically includes:

[0167] S4-1. The spectrum enhancement network model is trained using the gas sensor array data selected by the variable density sparse feature and the target label to obtain the trained spectrum enhancement network model.

[0168] S4-2. Input the gas sensor array data selected by the variable density sparse feature into the trained spectrum enhancement network model to obtain the predicted features used to characterize the gas type and concentration as the gas identification result.

[0169] In this embodiment, the data of the gas sensor array with variable density sparse feature selection and the corresponding labels are input into the constructed spectrum enhancement network model. The model is optimized by adjusting relevant parameters and the network model is fully converged through training iterations. Finally, the validation data of the single sparse feature selection completed in the S2 stage is input into the trained network model to predict the gas type and concentration. The prediction effect of the model is evaluated by relevant quantitative evaluation indicators.

[0170] The relevant quantitative evaluation indicators used in this embodiment specifically include the following:

[0171] Define success rate, precision, recall, Kappa coefficient, mean absolute error (MAE), and coefficient of determination (R²). 2 As performance evaluation metrics for classification and regression tasks, the number of parameters (params) and the number of floating-point operations per second (FLOPS) are used as evaluation metrics for model size and computational requirements. The formulas for these evaluation metrics are as follows:

[0172]

[0173]

[0174] Where n represents the total number of samples involved in the calculation, i represents the index number of the current sample, and y true It is the true value, y pred P is the predicted value, P0 is the observed consistency, and P... e It is based on the expected consistency under random prediction, and y'true is the average of the true values.

[0175] Based on the above, the specific steps for evaluating the model in this embodiment are as follows:

[0176] F1: Adjust relevant parameters and use training optimization and gradient descent algorithms to reduce the network model loss and increase relevant indicators until the network fully converges.

[0177] F2. The validation set data is preprocessed and variable density sparse feature selection is performed. The data is then input into the trained network model and the gas type and concentration are predicted through forward computation. The output is two tensors representing the predicted features of the gas type and concentration as the gas identification result.

[0178] F3. Calculate Accuracy, Precision, Recall, Kappa, MAE, and R. 2 Evaluation metrics such as Params and FLOPS.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0180] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0181] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0182] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0183] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A gas recognition method based on sparse feature selection and spectrum enhancement network, characterized in that, Specifically comprising the following steps: S1, data preprocessing is performed on the gas sensor array data set to obtain preprocessed gas sensor array data; S2, sparse feature selection is performed on the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels; S3, a spectrum enhancement network model is constructed; S4, the spectrum enhancement network model is used to obtain a gas identification result according to the variable density sparse feature selected gas sensor array data and the target labels. 2.The gas recognition method based on the sparse feature selection and the spectrum enhancement network according to claim 1, characterized in that, Data preprocessing is performed on the gas sensor array data set to obtain preprocessed gas sensor array data, which comprises: S1-1, obtaining sensor array data of a mixed gas and corresponding gas concentration and type labels as a gas sensor array data set; S1-2, sequentially performing normalization processing and linear interpolation processing on the gas sensor array data set to obtain preprocessed gas sensor array data. 3.The gas recognition method based on the sparse feature selection and the spectrum enhancement network according to claim 2, characterized in that, Sparse feature selection is performed on the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels, which comprises: S2-1, dividing the preprocessed gas sensor array data based on a Monte Carlo cross-validation method to obtain divided data of the preprocessed gas sensor array data; S2-2, using the divided data of the preprocessed gas sensor array data to obtain gas instantaneous energy distribution features of the gas sensor array data; S2-3, performing sparse feature selection on the gas instantaneous energy distribution features of the gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels. 4.The gas recognition method based on the sparse feature selection and the spectrum enhancement network according to claim 3, characterized in that, Using the divided data of the preprocessed gas sensor array data to obtain gas instantaneous energy distribution features of the gas sensor array data comprises: S2-2-1, performing complete ensemble empirical mode decomposition on the divided data of the preprocessed gas sensor array data to obtain a set of gas intrinsic mode functions; S2-2-2, performing Hilbert transform on the set of gas intrinsic mode functions to obtain gas instantaneous energy distribution features of the gas sensor array data.

5. The gas identification method based on sparse feature selection and spectrum enhancement network according to claim 3, characterized in that, Performing sparse feature selection on the gas instantaneous energy distribution features of the gas sensor array data to obtain variable density sparse feature selected gas sensor array data and target labels comprises: S2-3-1, using the gas instantaneous energy distribution features of the gas sensor array data to obtain feature energy guided sparse selection probability distribution; S2-3-2, constructing a variable density sparse feature selection matrix according to the feature energy guided sparse selection probability distribution; S2-3-3, using the variable density sparse feature selection matrix to perform variable density sparse feature selection and enhancement on the divided data of the preprocessed gas sensor array data to obtain variable density sparse feature selected gas sensor array data; S2-3-4, obtaining corresponding labels as target labels according to the variable density sparse feature selected gas sensor array data. 6.The gas recognition method based on the sparse feature selection and the spectrum enhancement network according to claim 1, characterized in that, The spectrum enhancement network model is stacked by a spectrum enhancement multi-scale TCN feature extraction layer, a linear projection layer, N FFT Fourier encoding layers, and a feature decoding layer with a double-branch structure. The spectrum enhancement multi-scale TCN feature extraction layer is composed of M TCN time sequence convolution networks and a spectrum enhancement layer. The Fourier encoding layer comprises a fast Fourier transform layer, layer normalization, a residual connection layer, and a feedforward neural network; The feature decoding layer of the double-branch structure comprises a plurality of fully connected layers and Dropout layers, and is configured with a SiLU activation function; wherein N FFT is a hyper-parameter to control the number of layers of the Fourier encoder, M TCN is a hyper-parameter to control the number of temporal convolution networks with different kernel sizes, each of which comprises a causal convolution layer and an inflated causal convolution layer, and is equipped with a batch normalization layer and a ReLU activation function, and each of which is inserted with a number of spectral enhancement layers.

7. The gas identification method based on sparse feature selection and spectrum enhancement network according to claim 6, characterized in that, The construction of the spectrum enhancement layer comprises: Performing adaptive average pooling operation on the one-dimensional feature sequence of the gas sensor array to obtain a feature sequence with a uniform sequence length; According to the frequency index set corresponding to the uniform sequence length, a frequency index set is obtained. According to the feature sequence with the uniform sequence length, a one-dimensional discrete cosine transform is performed to obtain a frequency domain feature value; Using the learnable weight parameters in the spectrum domain to perform adaptive weighting on the frequency domain feature value based on the frequency index set to obtain an adaptively weighted and fused spectrum feature; According to the feature sequence with the uniform sequence length and the adaptively weighted and fused spectrum feature, a spectrum enhancement layer is constructed by using a fully connected layer and an activation function to obtain a spectrum enhancement feature representation. Wherein, the one-dimensional feature sequence of the gas sensor array is a one-dimensional feature sequence of the gas sensor array extracted by a causal convolution layer or an expanded causal convolution layer of a time series convolution network. 8.The gas recognition method based on the sparse feature selection and the spectrum enhancement network according to claim 1, characterized in that, According to the variable-density sparse feature-selected gas sensor array data and the target label, the spectrum enhancement network model is used to obtain a gas recognition result, which comprises: S4-1, using the variable-density sparse feature-selected gas sensor array data and the target label to train the spectrum enhancement network model to obtain a trained spectrum enhancement network model; S4-2, inputting the variable-density sparse feature-selected gas sensor array data into the trained spectrum enhancement network model to obtain a predicted feature representing gas species and concentration as a gas recognition result.

Images