DNN-Aided MIMO Detection for Varying Wireless Channels

The patent applies preliminary action by performing offline machine learning training to determine optimal detection parameters before actual data reception. The DNN-based preprocessor is trained in advance using channel state information and reference signals to learn channel characteristics, so that when real data arrives, the detector can immediately use pre-learned parameters without complex real-time computation, thus achieving high precision with manageable complexity.

The patent introduces a DNN-based preprocessor as an intermediary between the received signal and the MIMO detector. This preprocessor takes channel state information and reference signals as input, processes them through machine learning to extract relevant features and determine optimal detection parameters, then feeds these processed parameters to the detector. This intermediary structure separates the complex learning task from the detection task, improving both precision and complexity management.

The patent uses preliminary machine learning training to pre-determine detection parameters that are optimized for specific channel conditions. When the number of antennas and users increases, the system can leverage the pre-learned parameters from the DNN preprocessor that were trained to handle such scenarios, maintaining detection reliability without being overwhelmed by the increased system complexity during actual operation.

The patent dynamically adjusts detection parameters based on channel conditions through machine learning. The DNN-based preprocessor learns optimal detection parameters (such as equalization filters, detection thresholds, and MIMO detection settings) that are adapted to the specific channel state, number of antennas, and user configuration. This parameter adaptation allows the system to maintain high detection reliability even as the number of antennas and users increases by using parameters specifically optimized for each scenario.

The patent resolves the time-precision tradeoff by performing the computationally intensive machine learning training in advance, during an offline phase. The DNN-based preprocessor is trained using channel state information and reference signals to learn optimal detection parameters before actual data reception begins. During real-time operation, the system only needs to apply the pre-learned parameters, which requires minimal processing time, thus achieving high detection precision without significant processing time penalty.

The patent segments the detection process into two distinct phases: an offline training phase where the DNN preprocessor learns optimal parameters using machine learning, and an online detection phase where pre-learned parameters are applied. This segmentation allows the computationally intensive learning process to occur when processing time is less critical (offline), while the time-sensitive detection operation uses the pre-computed parameters, thereby achieving high precision with minimal real-time processing overhead.