InertiEAR implementation

Data Collector:

We implement an Android App to collect data, the app is called SpyApp.

The source code of app is https://github.com/EdisonE3/IMU-Collect-App

The code of this app we are put into spyapp.zip.

The data we used to train is placed in the raw directory.

Importing this project into Android Studio or IDEA, then you can compile it.

After that, you can install this app on an android mobile phone and use it collect data.

Open the app, click the right low icon into collect mode.

Then you need to push the button test to collect data.

The data format is shown as below:

103094561358952,0.32542386651039124,0.1782652884721756,10.971092224121094
103094561372962,0.26081767678260803,0.19980068504810333,10.643275260925293
103094562586243,0.14835500717163086,0.23808585107326508,10.059426307678223
...

First column is time stamp, the following three columns are data collected from x, y and z axises from the mobile phone.

Model Training

• After data collect, use mian function in read_data.py

    out_dir_path = "files_train/signal_data_new"
    for file_name in os.listdir(in_dir_path):
        if file_name.count("acc"):
            acc_file = file_name
            gyr_file = file_name.replace("acc", "gyr")
            label = int(file_name.replace(".txt", "").split("_")[-2])
            print(acc_file, gyr_file, label)
            out_dir_path_i = out_dir_path + "/files_" + str(label)
            if not os.path.isdir(out_dir_path_i):
                os.mkdir(out_dir_path_i)
            data_processing(acc_path=in_dir_path + "/" + acc_file, gyr_path=in_dir_path + "/" + gyr_file,
                            file_directory=out_dir_path_i + "/", label=label)

where in_dir_path denotes the path where the original data exist and out_dir_path denotes the file in .npy

• Run the main function in Main.py with modified training model can be defined personally like myModel = SENet()

    myModel = torch.load("model_good/new/mobile_net.pth")
    myModel = myModel.to(device)

    # Training Model
    num_epochs = 20
    training(myModel, train_dl, val_dl, num_epochs)

    # Inference
    correlation_matrix = inference(myModel, val_dl, is_correlation=True)
    print(correlation_matrix)

    torch.save(myModel,"model/mobile_net.pth")

Num of epoch recommended to less 20 epochs for SENet.py to avoid overffiting.

Result Display

An implemented function segmentation in read_data.py can be used to display the segmentation process with input is_plot = True. This function should called first by estiblish a segmentation handler as h_seg = read_data.segmentation_handle(acc_xyz, gyr_xyz, acc_t, gyr_t should be , Fs = 400) where the acc_xyz, gyr_xyz,acc_t,gyr_t should be read by read_data.signal_read(path) where path is the cellphone sampled IMU data in txt format.

Reference

[1] https://github.com/Mayurji/Image-Classification-PyTorch/blob/main/DenseNet.py [2] https://github.com/soerenab/AudioMNIST