Neural Network Bit Recovery for FDMA-Based Wide-Area Ambient-IoT Networks

Ambient-IoT is gaining attention as a large-scale IoT technology for asset inventory management due to its low deployment costs and wide coverage. However, the low performance of passive tags causes a 'performance cliff' where collisions and bit errors increase simultaneously as the number of connected targets grows to thousands.

In existing Slotted ALOHA schemes, signals collide when multiple tags select the same time slot. Even without collisions, tags without FEC experience bit errors in poor backscatter channels; with a BER of 10^-2, the probability of successfully receiving a 256-bit ID is only about 7%.

To address this, Samsung applied Miller Code-based subcarrier FDMA. When tags use different Miller codes, signals are separated in the frequency domain, and tags randomly select not only time slots but also Miller codes, expanding the 1D contention space to 2D.

• 1-Miller, 4-Miller, 8-Miller, and 16-Miller signals use different frequency bands.
• If four Miller codes are available, the required inventory time slots can be reduced by 75%.
• This alleviates the bottleneck that required over 46,000 time slots to process 14,000 tags.

Secondly, a Discrete Hopfield Neural Network (DHNN) is applied at the base station to recover damaged IDs. Leveraging the fact that the base station already holds a valid Asset ID database, the received ID containing errors is treated as a pattern recognition problem rather than a communication error.

The DHNN is a recurrent network composed of fully connected neurons. After inputting the damaged ID as the initial state, the neuron states are iteratively updated, and the network converges to the closest state among the stored valid ID patterns, thereby recovering the bits.