Exploration for a BP-ANN model for gas identification and concentration measurement with an ultrasonically radiated catalytic combustion gas sensor

The ultrasonic radiation method provides a new solution to the single sensor based gas analysis. But it has been unknown whether the artificial neural network (ANN) can be effectively applied in the gas analysis with an ultrasonically radiated single gas sensor and how to apply. In this work, the BP-ANN model which can effectively implement the gas identification and concentration measurement with an ultrasonically radiated catalytic combustion gas sensor is explored, and a BP-ANN model with prominent performance in the gas identification and concentration measurement, named GWO-DHBP (double hidden layer BP), is found. Its feature set is designed with the assistance of the minimal redundancy maximal relevance (MRMR) method, and its initial weights and biases are optimized by the grey wolf optimization (GWO). The results show that the model has quite good gas recognition accuracy (97.3%) and small gas concentration measurement error (5.79%) in the gas concentration range of 2%-20%LEL (LEL=Lower Explosive Limit), with a faster convergence speed than the single-hidden-layer and Elman neural networks models with the GWO. The GWO is employed to overcome the BP-ANN's drawbacks such as easily falling into local minimum, slow convergence and poor generalization. It is demonstrated that the GWO-DHBP model is a promising algorithm for the gas identification and concentration measurement with the ultrasonically radiated catalytic combustion gas sensor, and a good feature vector may be achieved by using the experience, MRMR and the neural network which is going to be employed in the modeling.

Automated summary

This work explores the application of the BP-ANN model in gas analysis using an ultrasonically radiated catalytic combustion gas sensor. It identifies a model called GWO-DHBP with excellent performance in gas identification and concentration measurement. The model shows high accuracy in gas recognition and low measurement error, with a faster convergence speed compared to other neural networks models.