Abstract
In order to separate multiple random fault source signals adaptively only from mixed vibration measurement signals of mechanical system in the situation of time-varying source signals and mixture system, two adaptive blind identification and separation methods are proposed to solve this kind of problem. One is based on recursive least squares (RLS) algorithm and another is based on recursive EASI algorithm. The simulation results show that both of these methods can separate source signals from the mixed signal in the situation of time-varying source signals and mixture system very well.
Introduction
In mechanical fault diagnosis, engineers monitor vibration signals to identify failing components like bearings or gears. However, a single sensor rarely picks up a "clean" signal; instead, it receives a complex mixture of vibrations from various parts of the machine, often distorted by the structure's geometry (multipath effects). This is known as Blind Source Separation (BSS).
The major challenge addressed here is non-stationarity. In real-world operations, both the source signals (the faults themselves) and the mixing system (the machine's physical state) are often time-varying. Traditional BSS algorithms typically assume a static environment. If a machine speeds up, slows down, or its structural response changes, static algorithms fail to track and separate the signals accurately. This paper introduces two adaptive methods—Recursive Least Squares (RLS) and Recursive Equivariant Adaptive Separation via Independence (R-EASI)—to solve this "cocktail party problem" for industrial machinery.
Methodology
The authors propose two adaptive identification and separation methods designed to track changes over time. The general framework involves a time-varying mixing matrix $A(t)$ and a separation matrix $W(t)$ that must adapt to recover the original source signals $s(t)$.
- Recursive Least Squares (RLS) based BSS: This method utilizes a natural gradient algorithm to optimize the separation matrix in real-time. Unlike "batch" processing (which looks at a whole chunk of data at once), RLS updates its parameters at every time step, allowing it to "follow" the evolution of the system.
- Recursive EASI (R-EASI): The Equivariant Adaptive Separation via Independence (EASI) algorithm is a staple in signal processing. The authors enhance it by introducing a recursive reference point ($W_0$). By calculating a baseline separation matrix from a small window of recent samples, the algorithm avoids large accidental errors and improves tracking accuracy during the "tracking phase" of the signal.
By using recursive updates, the algorithm treats the "unmixing" process as a dynamic control problem rather than a static optimization problem. In R-EASI, the authors realized that standard adaptive filters can be jittery; using a baseline matrix ($W_0$) derived from recent observations provides a stabilizing influence.
Experimental Setup
The study relies on simulated vibration data to validate the algorithms. The simulation environment included three distinct source signals: a rectangular wave, a triangle wave, and a sine wave. To simulate real-world complexity, the sine wave's amplitude was modulated over time, and the mixing matrix (representing the physical system) had elements that decayed exponentially.
Results and Discussion
The simulation results demonstrated that both RLS and R-EASI could effectively recover the original waveforms from a messy, time-varying mixture. Visual comparisons between the original source signals and the separated output signals show high fidelity in waveform recovery.
Key findings include:
- Tracking Ability: The algorithms successfully tracked the amplitude modulation of the source signals and the shifting parameters of the mixing matrix without losing the "lock" on the individual sources.
- R-EASI Advantage: The R-EASI method specifically showed an improved ability to maintain stability and accuracy once it entered the tracking phase compared to standard adaptive approaches.
- Independence is Key: The methods rely on the statistical independence of the fault sources. As long as the faults (e.g., a bearing crack vs. a gear tooth chip) are physically unrelated, their signals can be untangled even if their frequencies overlap.
Conclusion
While the simulations are promising, ensuring the optimization method converges quickly and achieves higher precision remains a challenge. Future work will focus on performance in extremely high-noise industrial environments and reducing the computational overhead of recursive updates to enable deployment on edge computing devices mounted directly on machinery.
Summary updated