Rapid identification of multi-component hazardous gases in mines based on RDA-optimized 1D-CNN-BiLSTM

Rapid and accurate identification of multi-component hazardous gases in mines is the core premise for achieving early warning of hazardous gas leakage in mines. Traditional manual feature extraction methods can only reflect local discrete information in the response process and have low identification performance in complex scenarios such as multi-component gas leakage in mines. Existing gas identification algorithms based on machine learning and deep learning can automatically extract deep features in the spatiotemporal dimensions from gas sensor response data, but they need to wait until the sensor response reaches a steady state and cannot meet the demand for rapid gas identification. To address these problems, a hybrid neural network model of a One-Dimensional Convolutional Neural Network-Bidirectional Long Short-Term Memory Network (1D-CNN-BiLSTM) optimized by the Red Deer Algorithm (RDA) was proposed. The model extracted local transient features of gas responses through 1D-CNN and used BiLSTM to characterize the dependency relationships of long time-series data. It was able to perform end-to-end learning on gas sensor response data and avoid the subjectivity and limitations of manual feature extraction. RDA was introduced to adaptively optimize the core hyperparameters of the model, thereby improving model performance. The experimental results showed that the optimization efficiency and stability of RDA were better than those of the traditional Particle Swarm Optimization (PSO) algorithm and Genetic Algorithm (GA). The 1D-CNN-BiLSTM model was able to effectively extract gas category features with strong discriminative ability from gas sensor response data, and its identification accuracy for single gases and binary mixed gases was higher than that for ternary mixed gases. The identification accuracy of the proposed model reached 96.43%, which was better than those of traditional machine learning models and single-structure deep learning models. The model only used the gas sensor response data in the first 10 s after gas injection to achieve high-precision gas identification, balancing real-time identification and accuracy.