Introduction
Traditional signal processing systems typically rely on a fixed set of algorithms (SPAs) and static parameters, such as window sizes or sampling rates. While effective in controlled laboratory settings, these "one-size-fits-all" approaches often fail in complex, real-world environments where signals mask one another, noise levels fluctuate, and sources behave unpredictably. For instance, parameters optimized for a steady-state signal may miss impulsive events like a door slam, while high frequency resolution often comes at the cost of temporal precision.
The IPUS (Integrated Processing and Understanding of Signals) architecture addresses this by treating signal processing as an active, bidirectional search. Instead of seeking a single static configuration, the system dynamically reconfigures its low-level algorithms based on high-level interpretation logic, effectively changing its mathematical "lens" to resolve distortions.
The IPUS Architecture
The IPUS architecture is built upon the RESUN control framework and operates as an iterative loop designed to bridge the gap between signal processing theory and symbolic interpretation. The architecture can be viewed both as a generic framework and as an instantiated system for specific domains, such as acoustic signal processing.
The core process consists of four primary stages:
- Discrepancy Detection: The system monitors SPA outputs for inconsistencies against object models, environmental constraints, or the outputs of concurrent algorithms.
- Diagnosis: Using a formal theory of signal processing (e.g., Fourier theory), the system identifies the cause of a discrepancy. It employs "distortion operators" to hypothesize why a signal might be missing or blurred (e.g., insufficient frequency resolution).
- Signal Reprocessing: Once a distortion is diagnosed, IPUS generates a plan to re-run algorithms with adjusted parameters to eliminate the error.
- Differential Diagnosis: In cases of ambiguity between competing hypotheses, the system identifies distinguishing features and triggers targeted reprocessing of raw data to find specific evidence.
Evaluation and Results
The architecture was evaluated using a sound understanding testbed featuring a complex audio stream. The scenario included a constant hum from a hairdryer, an overlapping telephone ring, and a sudden door slam.
The system demonstrated several key capabilities in this testbed:
- Resolving Overlapping Signals: Initially, the hairdryer and phone ring appeared as a single blurred entity. IPUS detected a violation of its noise models, diagnosed a resolution issue, and reprocessed the data with a larger FFT size to successfully separate the sources.
- Capturing Transient Events: The system initially missed the "door slam" due to coarse time-decimation. By detecting a fault between the time-domain energy tracker and frequency-domain output, it reconfigured itself to capture the impulsive event.
- Pruning Ambiguity: When unable to distinguish a doorbell from a phone ring, the system used differential diagnosis to "zoom in" on a specific 2200 Hz component, providing the evidence needed for a correct classification.
Conclusion
IPUS represents a significant shift from passive to active signal processing. By leveraging formal signal processing theory as a diagnostic tool, the system can navigate the combinatorial explosion of possible parameter settings through selective, hypothesis-driven reprocessing. While future work remains to address real-time constraints in extremely noisy environments, IPUS provides a robust framework for integrated perception in complex domains.
Summary updated