Wearable-Activity-Classifier

Wearable Activity Classifier - Musa Khan

Task

Classify a 100-step, one-feature sensor sequence into Stationary, Walking, or Running.

Model selected

• Architecture: CNN+LSTM
• Input shape: (100, 1)[cite: 1]
• Output classes: 3[cite: 1]
• Parameters: 6,450 (Check your Colab table for the exact number)

Training data

Synthetic signals generated in the class notebook. The dataset was designed for teaching and is not a real wearable benchmark[cite: 1].

Evaluation

• Test accuracy: 97.5% (Check your Colab table for the exact number)
• Training time in our run: 45 seconds (Check your Colab table for the exact number)

Why we selected this model

We selected the CNN+LSTM hybrid because it combines the strengths of both architectures[cite: 1]. The Conv1D layer efficiently extracts short, localized temporal patterns (like the shape of a single step) and condenses the sequence length[cite: 1]. The LSTM then processes this resulting feature sequence to model long-range transitions over time, resulting in the most robust performance[cite: 1].

Limitations

• Synthetic, simplified data[cite: 1]
• One sensor feature only[cite: 1]
• No testing across real users/devices[cite: 1]
• Not intended for health, safety, or production use[cite: 1]

Team learning note

We learned that using MaxPooling1D between the CNN and LSTM layers is critical because it reduces the sequence length[cite: 1]. This condenses the strong features and prevents the LSTM from having to unroll over too many time steps.