Fig. 1. Image encoding process of Markov transition field

Fig. 2. Basic modules of MobileNetV2

Fig. 3. Impact of noise on MTF images. (a) Original signal and corresponding denoised signal; (b) MTF image of original signal; (c) MTF image of denoised signal

Fig. 4. Impact of number of quantile bins on MTF images. (a) Original knocking signal; (b) MTF image of knocking signal with Q=2; (c) MTF image of knocking signal with Q=5; (d) MTF image of knocking signal with Q=12

Fig. 5. Global and local features of MTF images. (a) Experimental signal; (b) MTF image corresponding to experimental signal

Fig. 6. Six types of signals and their corresponding MTF images. (a1)‒(f1) Background and signals of digging, knocking, watering, shaking, and walking; (a2)‒(f2) corresponding MTF images

Fig. 7. Training loss and validation accuracy of four different models. (a) Training loss; (b) validation accuracy

Fig. 8. Transfer learning

Fig. 9. Training loss and validation accuracy of MobileNetV2. (a) Training loss; (b) validation accuracy