Equipment State Detection With Bit-Reduced CNN Processing

The patent replaces the mechanical/time-series processing approach (RNN/LSTM) with a visual/image-based processing approach (CNN). By converting sensor signals into image representations, the system leverages the efficiency of CNN architectures which are optimized for spatial pattern recognition, thereby reducing computational complexity while maintaining detection accuracy.

The patent transforms the input data format from time-series sensor signals to image representations. This parameter change in data structure enables the use of CNN instead of RNN/LSTM, fundamentally changing the processing paradigm from sequential temporal analysis to parallel spatial analysis, which reduces device complexity.

The patent substitutes the computationally intensive Transformer architecture with a more efficient CNN approach. By representing sensor data as images and using CNN's localized filtering and parameter sharing mechanisms, the system achieves comparable detection accuracy with significantly reduced operational overhead and power consumption.

The patent segments the analysis process into two stages: initial screening with low-resolution processing, and detailed analysis with high-resolution processing only for regions of interest. This segmentation reduces the total operation amount and power consumption by avoiding full high-resolution processing across the entire dataset.

The patent divides the image processing into two resolution levels. The first CNN processes low-resolution images to identify potential abnormal regions, and the second CNN processes only those specific regions at high resolution. This segmented approach maintains detection accuracy for critical areas while reducing overall power consumption by limiting high-resolution processing to necessary regions only.

The patent applies different processing qualities to different regions of the image. Critical regions showing abnormal patterns receive high-resolution processing, while normal regions are processed at low resolution or skipped entirely. This local quality differentiation optimizes the balance between detection accuracy and power consumption.

The patent implements a two-stage processing pipeline where the first stage uses reduced bit-depth processing for initial screening, and the second stage uses full precision for detailed analysis of identified regions. This segmentation allows the system to benefit from lower power consumption in the majority of processing steps while preserving accuracy where it matters most.

The patent applies reduced precision processing partially - only to the extent necessary for initial detection and region identification. Full precision is applied excessively only to the specific regions that require detailed analysis, rather than uniformly across all data. This partial application of high precision optimizes the trade-off between power consumption and accuracy.