Multichannel convolution neural network for gas mixture classification

Abstract

Concomitant with people beginning to understand their legal rights or entitlement to complain, complaints of offensive odors and smell pollution have increased significantly. Consequently, monitoring gases and identifying their types and causes in real time has become a critical issue in the modern world. In particular, toxic gases that may be generated at industrial sites or odors in daily life consist of hybrid gases made up of various chemicals. Understanding the types and characteristics of these mixed gases is an important issue in many areas. However, mixed gas classification is challenging because the gas sensor arrays for mixed gases must process complex nonlinear high-dimensional data. In addition, obtaining sufficient training data is expensive. To overcome these challenges, this paper proposes a novel method for mixed gas classification based on analogous image representations with multiple sensor-specific channels and a convolutional neural network (CNN) classifier. The proposed method maps a gas sensor array into a multichannel image with data augmentation, and then utilizes a CNN for feature extraction from such images. The proposed method was validated using public mixture gas data from the UCI machine learning repository and real laboratory experiments. The experimental results indicate that it outperforms the existing classification approaches in terms of the balanced accuracy and weighted F1 scores. Additionally, we evaluated the performance of the proposed method in various experimental settings in terms of data representation, data augmentation, and parameter initialization, so that practitioners can easily apply it to artificial olfactory systems.

Appendices

Appendix A. The UCI dataset: data characteristics

• I. Sensor information to measure gas mixture

  The UCI data were collected using eight sensors, as shown in Table 10. In particular, the MOX gas sensor value used for data measurement varies according to the change in the resistance value of the semiconductor oxide film, at which time the oxide film of the semiconductor reacts to the gas molecule. Therefore, the response value of mixed gases cannot simply be expressed as a linear combination of the response values of individual gases and, to accurately predict the type and concentration of the desired gases, it must be able to reflect the characteristics inherent in mixed gases described earlier.

• II. Details of data collection and experiments

  The UCI dataset utilizes three types of gases: carbon monoxide (CO), methane, and ethylene. The experimental environment had two gas outlets: one outlet emitted 2500 ppm of ethylene, while the other outlet emitted 1000 ppm of methylene and 4000 ppm of CO. Each emission speed was varied, and the released gases were sprayed through turbines, making them a mixture of varying concentrations when the wind-blown gases touched the sensors. The concentrations of N, L, M, and H for each gas are listed in Table 11. The actual generated gas mixture samples were classified into five classes, as shown in Table 12.

Table 10 Sensor details and representation of UCI data

Table 11 Concentration levels in the UCI dataset

Table 12 Class information in the UCI dataset

Appendix B. Real data: data characteristics

Table 13 Sensor details and representation of real data

Table 14 Experimental setting of real data

Table 15 Concentration levels in the real data

• I. Sensor information to measure gas mixture

  The data used in this study were collected using a total of 11 gas sensor arrays, as shown in Table 13 as data from the Taesung Environmental Research Institute. There were two MOX type, one NDIR type, and eight ELECTRONICHEMICAL type sensors.

  The MOX sensor is based on the conductivity change of the gas-sensitive MOX semiconductor layer. The NDIR sensor recognizes gas molecules using infrared light. When infrared rays are absorbed by gas molecules, they vibrate the gas molecules at specific wavelengths. The electronic chemical sensor makes measurements by oxidizing or reducing the gas it wants to detect when it enters the internal sensor.

• II. Details of data collection and experiments

  The data from the Taesung Environmental Research Institute were created by varying the injection volumes of three types of standard gases: dimethyl sulfide, butyl acetate, and toluene. Standard gas was prepared at each concentration of 10,000 ppm using the following formula:

  $$\begin{aligned} V_i[\mu l] = \frac{C_t[ppm] \times V_t[L] \times w_m[g/mol]}{22.4 \times 1000 \times d[g/ml]} \times \frac{\textit{p}[\%]}{100}. \end{aligned}$$

  $V_i$ represents the amount of injection gas (injection volume), $C_t$ represents the target concentration (target concentration), $V_t$ represents the target volume (Target Volume), $w_m$ indicates the atomic volume of the injection gas (density), and p indicates the purity of the injection gas (purity). The gas was injected to match the concentration and volume of the target mixture, and the experiment was conducted, as shown in Table 14.

  A mixture of 1ml=0.001l of standard gas generated by the above process and 10l of nitrogen can produce 1ppm of mixed gas, and several standard gases can be used to produce mixed gases with different proportions and concentrations.

  A total of 100 samples were generated and classified into four classes according to the ratio of the injection amount, with the characteristics and numbers of each class equal to those in Table 15. The numbers in parentheses indicate the amount of standard gas injected during the sample generation process. For example, class 1 is a specimen injected with 1000 ppm of dimethyl sulfide and 1000 ppm of toluene. In this study, we focused on the existence of a specific chemical in a gas mixture rather than its ratio or amount. Therefore, in the case of class 4, diverse combinations of concentrations were treated as the same class.

Appendix C. Details of model architecture and learning strategy

• I. Model architectures

  There are several major differences between ResNet models for ImageNet classification and Multichannel ResNet models for gas mixture classification. First, the input size and normalization parameters differ. The ImageNet input images were cropped to a size of $224\times 224$ with normalization with three channels: mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225]. The proposed model uses a larger number of channels. We use a larger resizing to $256\times 256$ and the same normalization for all channels, such as mean=$[0.5, \cdots , 0.5]$ and std=$[0.5, \cdots , 0.5]$.

  The first input layer is changed because of the larger input size, which aggregates the multichannel input to the designated output dimension. The first convolutional layer uses a smaller kernel size and stride. Therefore, the proposed architecture does not reduce the input after the first convolutional layer. ResNet works as downsampling to reduce the image and computation. In our case, we used more expensive operations because of the small number of data and features. In addition, the last fully connected layer changed to the number of classes, from 1000 labels in ImageNet to four or five labels in our cases. This is a general modification of transfer learning. Therefore, we changed the major components in the model for our datasets and tasks.

• II. Training Parameters Setting

  Five-fold cross validation was used to avoid over-fitting. The models were trained with a batch size of 4, Adam optimizer, learning rate of $10^{-6}$ with weight decay of $10^{-8}$, and 50 epochs. The weighted cross-entropy loss function was used, and the best model was selected based on the minimum loss in the validation set.

Table 16 Comparison model: ResNet18 versus multichannel ResNet18

Table 17 Comparison model: ResNet50 versus multichannel ResNet50

Appendix D. The UCI dataset: Model Performances

• (I) Machine models with the Bag-of-Features representation

• (II) Comparison of Noise-added data augmentation

• (III) CNN models with the analogous image matrix representation

Table 18 The UCI dataset: performance metrics for BoF based models

Table 19 The UCI dataset: performance metrics for noisy augmented data

Table 20 The UCI dataset: performance metrics for analogous image matrix-based models

Appendix E. Real data: Model Performances

• (I) Machine models with the Bag-of-Features representation

• (II) Comparison of Noise-added data augmentation

• (III) CNN models with analogous image matrix representation

Table 21 Real data: Performance metrics for BoF-based models

Table 22 Real data: Performance metrics for noisy augmented data

Table 23 Real data: performance metrics for analogous image matrix-based models

Appendix F. The bag-of-features approach

The BoF approach represents complex objects with feature vectors of sub-objects. Originally, the main idea of the BoF approach is well-known as bag-of-words (BoW) and developed in document classification. The main idea is to define a code book and display it as a histogram, with each item being the number of code books that have occurred in the document. While ignoring word order information, BoW models are still very effective in capturing document information because they are good at navigating document code word frequency information (Zhang et al., 2010; Schütze et al., 2008) Also, BoF is a well-known dictionary-based feature extraction method for time-series classification (O’Hara & Draper 2011; Pancoast & Akbacak 2012; Baydogan et al., 2013). Furthermore, BoF can be easily applied when the observations have different sequence lengths (Tan et al., 2019).

Feature selection or feature extraction is a common approach to analyze high-dimensional data. We implemented BoF representation for multivariate time series. The BoF representation can represent the original data into lower-dimensional features. We had N samples of the gas sensor array in each dataset. In this study, the bag-of-features representation was the baseline for evaluating the proposed approach. compared to the image transformation, we used the individual observations in the gas sensor array as a feature:. ${\mathbf {X}} =\left[ {\mathbf {z}}_1\; {\mathbf {z}}_2\;\cdots \; {\mathbf {z}}_T\right]$, where ${\mathbf {z}}_j= (x_{1j}, \ldots , x_{Mj})^T$, $j=1,\ldots ,T$. Here, we use ${\mathbf {z}}_j$ as a sample for the BoF.

To build a codebook, ${\mathbf {z}}_j$ can be quantized using a clustering algorithm, hence the K-means clustering model g with $K=128$ was trained. K-means clustering generates K centroids, and each vector ${\mathbf {z}}_j$ is classified with the centroid index closest to it. In other words, $g ({\mathbf {z}}_j) = k$, where $j \in \left\{ 1, \cdots , T\right\}$ and $k \in \left\{ 1, \cdots , K\right\}$. Subsequently, we used the centroid index as a codebook for each sample. Each centroid can be seen as a “word” in the dictionary and each sample as a “document, ” which has a different set of words. We used the frequency ratio of each centroid index as a feature of the samples. In this approach, K is the size of the codebook, which can affect the classifier model, but the effect of the codebook size was not investigated in this research.

Fig. 13

Sample data plot

We used a sufficiently large K to represent each sample’s characteristics. Following the proposed approach, each sample was transformed into a K-dimensional vector. The original data $\mathbf{X}$ became a feature vector $\mathbf{f}$ with a size of K.

$$\begin{aligned} \mathbf{f} = \left[ f_k \mid f_k = \frac{\# (z \textit{ classified as cluster } k)}{T} \quad \textit{where } k \in \left\{ 1, \cdots , K\right\} \right] \end{aligned}$$

Through this process, we could extract K-dimensional features from the $M \times T$ matrix. Once features were selected by using the bag-of-features, we used logistic regression (LR), support vector machine (SVM), Naive Bayes (NB), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LGBM), and multilayer perceptron (MLP) to classify the classes of samples.