Abstract
This paper presents an online identification algorithm for Takagi-Sugeno-Kang (TSK) fuzzy models. The proposed algorithm identifies the structure of the premise part using distance metrics and obtains the parameters of the piecewise linear functions in the consequent part using recursive least squares (RLS). While most existing research considers only the input space for structure identification, the proposed algorithm incorporates both input and output spaces. This dual-space approach effectively excludes outliers during clustering. Furthermore, unlike many existing algorithms that use sample data points as fixed cluster centers—making them sensitive to noise—the proposed method dynamically updates centers, enhancing robustness. The model is obtained in a single pass without requiring extensive data storage.
Introduction
Identifying the structure of nonlinear systems in real-time is a cornerstone of modern control theory and signal processing. TSK fuzzy models are particularly effective for this task as they bridge the gap between the interpretability of fuzzy logic and the precision of local linear models. However, traditional online identification methods often struggle with two primary issues: vulnerability to outliers due to input-only clustering and high sensitivity to noise caused by static cluster center assignment. This research addresses these gaps by introducing a robust, one-pass learning framework that is resilient to sensor interference.
Methodology
The proposed identification process consists of two main phases: premise identification and consequent parameter learning.
Premise Identification (Structure Learning)
The "If" part of the fuzzy rules is determined through a novel clustering approach that utilizes Dual-Space Evaluation. By calculating distances in both the input and output spaces, the algorithm can distinguish between a genuine shift in system state and a transient outlier. A new rule is only created when a data point is significantly distant in both spaces. Additionally, the algorithm employs Dynamic Center Updates, where the center and width (standard deviation) of Gaussian membership functions are refined as new data arrives, effectively smoothing out the impact of initial measurement noise.
Consequent Identification (Parameter Learning)
The "Then" part of the TSK model consists of linear functions updated via Recursive Least Squares (RLS). This allows the model to adapt its local linear approximations on the fly, ensuring high accuracy without the need to store historical data.
Experimental Setup
The algorithm was rigorously tested using the Mackey-Glass (MG) chaotic time series, a standard benchmark for nonlinear system identification. To evaluate robustness, Gaussian noise was added to the MG data to simulate real-world sensor interference. This setup tests the model's ability to extract the underlying signal from a corrupted data stream.
Results and Discussion
The proposed algorithm demonstrated superior performance compared to established benchmarks such as DENFIS, eTS, and EfuNN.
- Accuracy and Efficiency: On the Mackey-Glass benchmark, the method achieved a Non-Dimensional Error Index (NDEI) of 0.0176 using only 50 rules. In contrast, the eTS model required 113 rules to reach an NDEI of 0.0954, and EfuNN required 1125 rules for an NDEI of 0.094. The efficiency gains are particularly noteworthy; while EfuNN achieves comparable accuracy, its requirement for over a thousand rules makes it impractical for real-time applications where memory and latency are critical.
- Outlier Rejection: A critical innovation is the validation of new clusters using the output space. By verifying if the corresponding output also deviates significantly, the algorithm avoids creating redundant or erroneous rules, maintaining a compact and interpretable model.
- Robustness: The dual-space clustering effectively ignored spikes and sensor glitches that typically lead to "rule explosion" in other systems. Even under severe noise conditions, the tracking performance remained remarkably stable, as shown in the comparison between the real function and the model's output.
Conclusion
The proposed one-pass, online evolving TSK fuzzy identification algorithm offers a robust and computationally efficient solution for real-time modeling. By integrating output-space information into the clustering process and dynamically updating rule parameters, it achieves high accuracy with a parsimonious rule base. Future work will explore automated hyperparameter tuning to further enhance the algorithm's autonomy in unknown environments.
Summary updated