Hui-Rang Hou

Detecting ignitable liquids (ILs) at fire scenes is essential for fire investigations. Electronic nose (e-nose) plays a crucial role in the detection of ILs due to its low cost and rapid response. However, the performance of most classification algorithms is often degraded when deployed across different e-noses devices. To address this issue, we first propose a sensor screening strategy for e-noses to remove sensors with apparently inconsistent signals, and then we put forward a novel domain adaptation algorithm called selective kernel feature alignment convolutional neural network (SKFA-CNN), which utilizes kernel embedding to facilitate the learning of domain-invariant features across e-noses in a reproducing kernel Hilbert space. SKFA-CNN aims to aggregate multi-scale information from multiple branches to enhance the comprehension and representation of e-nose signal features and to fuse multi-kernel features adaptively and nonlinearly. To evaluate the proposed method, five e-noses with identical sensor arrays were fabricated based on the screened sensors, and a dataset was built by collecting data on four types of ILs (i.e., ethanol, gasoline, diesel and diluent) using the self-designed e-noses. Experimental results demonstrate that this method outperforms current state-of-the-art IL detection techniques, achieving an average recognition accuracy of 94.76%, an F1-score of 94.82%, and a PAM of 90.64%.

Identification of ignitable liquids (ILs) plays an important role in fire investigation. Electronic noses (E-noses) have been used in the field of ILs identification due to their advantages of fast speed, low price and good portability. However, when we ran the same algorithm on different E-noses with the same sensor array (i.e. the same type and number of gas sensors) to identify the same ILs, we found that the identification results were different, mainly because of the poor consistency of the gas sensors used in the E-noses. To make the identification results of all E-noses essentially the same, each E-nose must be trained with a large amount of data individually, which would waste a lot of time. To address the issue of ILs identification across E-noses, we proposed a multi-branch adaptive feature fusion network (MAFF-Net) that can simultaneously process one-dimensional (1D) data directly output from E-noses and two-dimensional data converted from 1D data based on Gramian angular field. The MAFF-Net can be trained using one E-nose and then transplanted to another E-nose by fine-tuning its model parameters with a small training dataset. To verify the performance of the proposed MAFF-Net, three E-noses were designed and fabricated, and an ILs dataset was constructed based on the self-built E-noses. The experimental results show that the recognition accuracy of the MAFF-Net algorithm on E-nose No.1, No.2 and No.3 is 100.0%, 97.5% and 90% respectively, which is better than the five classical algorithms.

The identification of ignitable liquids at fire scenes is a crucial component of forensic investigations.

Electronic nose (e-nose) can carry out odor fingerprinting and identification, exhibiting several advantages such as speed, low cost and portability. The e-nose technology allows objective, accurate and real-time detection and recognition of ignitable liquids (ILs), which provide technical support for fire cause determination and fire scene reproduction. Metal oxide semiconductor (MOS) gas sensors have certain technical limitations, and multiple e-nose platforms assembled using the same sensor array will produce different output responses when sampling the same gas. Manual data labeling and repetitive training of models involve a huge workload. This study proposes a novel domain generalization method to mitigate this problem. During the training process, we assign pseudo-domain labels iteratively for the e-nose response signals obtained from multiple source domains by clustering. This approach avoids the situation where multiple potential domains can exist for the same e-nose. The highlights of our approach are as follows: 1) The training paradigm of traditional domain generalization methods that require domain labels is not used; 2) Pseudo-domain labels are generated for the samples according to each cycle iteration during training; 3) The feature extractor is trained to learn domain-invariant features by adversarial learning based on pseudo-domain labels; 4) The statistics of each hidden layer of the neural network are used to cluster the e-nose signals and generate pseudo-domain labels from them. Five e-noses, each with 16 identical sensor channels, are used to recognize four types of ILs. Our method achieves an average cross-platform recognition accuracy of 85.2%.

Detecting ignitable liquids (ILs) remaining at the fire scene is a critical task in fire investigation. Current detection methods mainly rely on large analytical instruments, which suffer from slow detection speeds and high costs. Electronic nose (e-nose) has been widely used in the fields of food testing, environmental monitoring, and disease diagnosis due to its advantages of fast detection speed and low cost. However, there has been limited research on the detection of ILs using e-nose technology. This study introduces a novel identification method called the channel-separated multiscale attentional convolutional neural network (CS-MA-CNN). The CS-MA-CNN utilizes a portion of the response time data from an e-nose to achieve rapid identification of ILs. The highlights of CS-MA-CNN are as follows: 1) CS-MA-CNN is a channel separation technique for one-dimensional CNNs that effectively captures the correlations between the temporal and channel dimensions of e-nose data. 2) A multi-scale attention module is applied to enhance important features in the main channel and temporal feature maps, and suppress the redundant features. 3) The fully connected-layer network is used for classifying the ILs labels. Recognition experiments were conducted on four types of ILs using a homemade portable e-nose. The results show that the CS-MA-CNN achieved an impressive classification accuracy of 94.53% with just 5 seconds of the e-nose response data.

Herein, an original algorithm, the model-free one-class open-style circular coding (OC-OCC) approach is proposed to identify the authenticity of Chinese liquors based on an electronic nose system. When measuring liquors with an e-nose, the liquor odors were transformed into the corresponding odor fingerprint, which is composed of the response curves of the gas sensor array. In the training stage, only the positive samples, i.e., the true liquors’ odor fingerprints were used to train algorithms. For feature extraction, the response curve of each gas sensor was averagely divided into 20 segments, followed by a median average of the response voltages in each segment to form the voltage features. A circular feature set is subsequently constructed using the voltage features of each segment, and the circular codes are generated by arranging features of different sensors in ascending order. In the identification stage, the testing samples with the same circular codes as the trained true (positive) liquor type are predicted as true samples, otherwise as the fake (negative) samples. The proposed method yields average accuracy of 98.7% and 96.1% for two public datasets, superior to three classical one-class classification algorithms, such as the one-class support vector machine (OC-SVM), isolation forest (IF), and the one-class k-nearest neighbor (OC-KNN).

Objective. Depression is a common chronic mental disorder characterized by high rates of prevalence, recurrence, suicide, and disability as well as heavy disease burden. An accurate diagnosis of depression is a prerequisite for treatment. However, existing questionnaire-based diagnostic methods are limited by the innate subjectivity of medical practitioners and subjects. In the search for a more objective diagnostic methods for depression, researchers have recently started to use deep learning approaches. Approach. In this work, a deep-learning network, named adaptively multi-time-window graph convolutional network (GCN) with long-short-term memory (LSTM) (i.e. AMGCN-L), is proposed. This network can automatically categorize depressed and non-depressed people by testing for the existence of inherent brain functional connectivity and spatiotemporal features contained in electroencephalogram (EEG) signals. AMGCN-L is mainly composed of two sub-networks: the first sub-network is an adaptive multi-time-window graph generation block with which adjacency matrices that contain brain functional connectivity on different time periods are adaptively designed. The second sub-network consists of GCN and LSTM, which are used to fully extract the innate spatial and temporal features of EEG signals, respectively. Main results. Two public datasets, namely the patient repository for EEG data and computational tools, and the multi-modal open dataset for mental-disorder analysis, were used to test the performance of the proposed network; the depression recognition accuracies achieved in both datasets (using tenfold cross-validation) were 90.38% and 90.57%, respectively. Significance. This work demonstrates that GCN and LSTM have eminent effects on spatial and temporal feature extraction, respectively, suggesting that the exploration of brain connectivity and the exploitation of spatiotemporal features benefit the detection of depression. Moreover, the proposed method provides effective support and supplement for the detection of clinical depression and later treatment procedures.