I've considered total on time of each sensor as featureto activity recognition, but it doesn't give me good accuracy, i'm working with MIT dataset

Dear Malihe Erfanmanesh,

I have some works associated with activity recognition using wearable sensor data. In my papers, I have achieved state-of-the-art results. I think you can find good ideas and directions to future works in these works:

"Human Activity Recognition Based on Wearable Sensor Data: A Standardization of the State-of-the-Art"

"Latent hypernet: Exploring all Layers from Convolutional Neural Networks"

"Novel Approaches to Human Activity Recognition based on Accelerometer Data"

Depending of dataset , CNN requires large dataset while ANN can do with smaller. Try NN again you need input vs target here .