A CNN-LSTM model for road surface recognition of electric balance vehicles

Abstract

In recent years, artificial intelligence has developed rapidly, and artificial intelligence equipment has greatly facilitated people's lives. During the operation of intelligent devices, accurate environmental perception is of great importance. To provide more accurate electric balance vehicle instructions, we collect the vibration signal data in different driving environments, and machine learning (ML) methods are adopted to extract and analyse the characteristics of the signal. Then, the classification of the road type is realized. In the study, a Bluetooth wireless acceleration-attitude sensor is used to collect acceleration, angular velocity and angle signal data of the balance vehicle in the X, Y and Z directions. Through analysis, the classification accuracy of using continuous time series data as learning samples is higher than that of using a single time step signal. In the classification based on continuous time series samples, three different models XGBoost, LSTM and 1D-ResNet, are used. XGBoost is applied on artificial feature extraction data and the other models are applied on original sequence data. The F1 scores of the three models are 0.80, 0.77 and 0.79, respectively. To mitigate overfitting, we propose a CNN-LSTM model, whose F1 score is 0.83, outperforming the previous three models.

Keywords

• Sequential feature extraction
• Pattern classification
• CNN-LSTM
• Difficult sample mining