Parallel Disturbance Observer

• Parallel Disturbance Observer is a robust control architecture that uses an inner-loop observer running parallel to the nominal controller to estimate and cancel disturbances.
• It employs a Q-filter to balance high-frequency noise mitigation with effective disturbance rejection, based on both continuous and discrete-time analyses.
• Widely used in robotics, RF systems, and FPGA-based controls, parallel DOB offers systematic stability, rigorous performance guarantees, and precise disturbance management.

A parallel disturbance observer (parallel DOB or parallel DOb) is a robust control architecture in which an inner-loop observer estimates the lumped effect of model uncertainty and external disturbances by running in parallel with the nominal plant or controller. The disturbance estimate is subtracted in feedforward from the control input, thereby providing broadband disturbance rejection without compromising the outer-loop performance controller structure. The parallel DOB paradigm is dominant in high-precision motion systems, precision servos, FPGA-based control in scientific instrumentation, high-power radio-frequency systems, and networked motion platforms. Its effectiveness is underpinned by systematic stability and performance guarantees under well-posed tuning constraints, with design rules grounded in both continuous- and discrete-time analyses (Sariyildiz et al., 2019, Sariyildiz et al., 2019, Sariyildiz, 2021, Seo et al., 25 May 2026, Kwon et al., 2022).

1. Core Architecture and Mathematical Formulation

The defining feature of the parallel DOB is its block-diagram structure, with two distinct parallel paths:

• The "performance" controller (e.g., PID, PD, PI) generates a nominal input $u_{des}(t)$ based on the error between reference $r(t)$ and measured output $y(t)$.
• In parallel, the DOB loop uses a nominal model $G_n(s)$ (continuous time) or $P_n(z)$ (discrete time) and filtered error signals to estimate the total lumped disturbance $d(t)$ acting on the plant.

The plant output is modeled as

$y(s) = G(s)[u(s) + d(s)]$

with $G(s)$ the true plant and $d(s)$ the total (unmeasured) disturbance input. The nominal model output is $y_n(s) = G_n(s) u(s)$.

The disturbance estimate is

$r(t)$0

where $r(t)$1 is a causal, low-pass "Q-filter" that regularizes high-frequency noise. The plant input is then

$r(t)$2

This structure ensures that, in the ideal case ($r(t)$3, perfect state knowledge), low-frequency disturbances are rejected up to the Q-filter cutoff (Sariyildiz et al., 2019).

When extended to multi-input, multi-output (MIMO) systems (e.g., robot manipulators), the same structure is preserved, with vector- or matrix-valued state, disturbance, and observer components (Sariyildiz et al., 2019).

2. Stability, Robustness, and Sensitivity Analysis

The closed-loop disturbance-to-output transfer function is

$r(t)$4

At low frequencies ($r(t)$5, $r(t)$6), $r(t)$7, yielding strong disturbance rejection. At high frequencies ($r(t)$8), no rejection occurs, avoiding excessive noise amplification (Sariyildiz et al., 2019, Sariyildiz et al., 2019).

Robustness and ultimate boundedness (UUB) of the disturbance estimate and closed-loop system can be rigorously established. For example, for robot manipulators with nominal inertia $r(t)$9, the error bound is (regulation or tracking)

$y(t)$0

where $y(t)$1 is the observer bandwidth, and $y(t)$2 and $y(t)$3 capture inertia mismatch and gravity bound (Sariyildiz et al., 2019).

In discrete time, design is constrained by a Bode-integral "waterbed" effect and explicit upper bounds on the normalized DOB bandwidth, ensuring stability and avoiding excessive noise sensitivity. Key constraints are

$y(t)$4

where $y(t)$5 encodes nominal-to-actual model mismatch, $y(t)$6 the low-frequency sensitivity, and $y(t)$7 the permissible sensitivity peak (Sariyildiz, 2021).

3. Q-Filter Design and Performance Trade-Offs

The Q-filter $y(t)$8 (or $y(t)$9 for digital systems) is crucial. The filter cut-off sets the disturbance rejection bandwidth: a higher cut-off allows faster disturbance estimation/rejection but at the cost of amplifying noise and reducing robustness to model uncertainty and sampling effects.

• For analog/continuous IO: typically, $G_n(s)$0 (first order) or a well-damped second-order form.
• Digital implementations frequently use a backward-Euler discretization with $G_n(s)$1 (Sariyildiz, 2021).

Optimal tuning places the Q-filter cut-off above the expected disturbance bandwidth but well below the frequency at which sensor noise or unmodeled plant dynamics become significant. Design rules empirically place $G_n(s)$2 at $G_n(s)$3–$G_n(s)$4 the outer-loop bandwidth for motion control, subject to hardware noise limits (Sariyildiz et al., 2019). In digital control, the normalized $G_n(s)$5 is generally restricted to $G_n(s)$6–$G_n(s)$7 for robust operation (Sariyildiz, 2021).

For high-power radio-frequency systems, a two-pole real Q-filter can be used to cancel plant zeros and further tailor bandwidth, as in LANSCE DTL SSPA control (Kwon et al., 2022).

4. Extensions and Adaptive/Parallel Estimation Structures

Parallel DOB schemes can be extended by incorporating adaptive and learning components. For example, in the Self-Learning DOB (SLDO), a fast "basic" nonlinear DOB provides a bias-prone first estimate while an interval Type-2 neuro-fuzzy system (T2NFS) branch learns and cancels residual bias for arbitrary time-varying disturbances. Adaptive update laws, derived via sliding mode control (SMC) theory and feedback-error learning, ensure that the composite estimator converges with robustness to measurement noise (Kayacan et al., 2021).

This "parallel" fusion of conventional DOB and data-driven/adaptive estimation demonstrates significantly improved tracking of non-stationary disturbances, with formal Lyapunov proofs ensuring overall stability and performance. Such composite architectures are becoming relevant as system complexity and operating environments increase in uncertainty and variability.

5. Implementation in Discrete-Time and FPGA-Based Systems

In embedded and high-speed digital signal processing environments, the parallel DOB is implemented in discrete time, with Q-filters realized as exponential moving averages and fixed-point arithmetic for computational efficiency. Digital DOB stabilization in motion control follows tight constraints on filter gain, order, and sampling rate to avoid instability and excessive noise amplification. Empirical guidelines set sampling rates $G_n(s)$8–$G_n(s)$9 the closed-loop bandwidth, with normalized DOB gain chosen below the derived robustness bounds (Sariyildiz, 2021).

FPGA-based parallel DOB servos for laser frequency stabilization employ a minimal-parameter (gain, bit-shift) Q-filter, enabling high sampling rates (e.g., 125 MHz) with extremely low latency (e.g., 40 ns), full sampling-rate disturbance rejection, and simple tuning (Seo et al., 25 May 2026). These implementations dispel the misconception that detailed frequency-domain plant identification is required for robust DOB synthesis: simple gain sweeps for Q-filter tuning are sufficient for optimal noise suppression in low-order systems.

6. Representative Applications and Case Studies

Parallel DOBs are deployed in a wide range of application domains:

Application domain	Representative implementation	Performance metrics/remarks
Robot manipulators, servos	Acceleration-based robust position control (Sariyildiz et al., 2019)	Asymptotic regulation, ultimate boundedness, noise/robustness trade-off.
High-fidelity networked control	Parallel DOB for time-delay/communication disturbance	Compensation of delay-induced disturbances (Sariyildiz et al., 2019).
RF power amplifier phase control	Parallel DOBC on LLRF FPGA, SSPA–cavity system (Kwon et al., 2022)	SSPA phase drift $P_n(z)$0, amplitude error $P_n(z)$1.
Laser frequency stabilization	PID + digital parallel DOB on FPGA (Seo et al., 25 May 2026)	16.9 dB broadband noise reduction below 40 kHz; factor $P_n(z)$2 improvement in short-term stability.
Sensorless converter, automotive	PWM converters, yaw-moment observer (Sariyildiz et al., 2019)	Sensorless regulation, disturbance suppression.

In advanced scenarios, multi-channel and composite learning-based DOBs enable robust compensation of structured and unstructured disturbances, as evidenced by simulation and experimental benchmarks for nonlinear oscillators and under time-varying noise (Kayacan et al., 2021).

7. Comparative Discussion and Practical Guidelines

Parallel DOBs offer distinct advantages over series DOBs in terms of direct feedforward cancellation, ease of Q-filter tuning, and modular controller integration. However, the need for adequate nominal model fidelity and careful bandwidth selection persists: excessive observer gain or inappropriate nominal plant matching can lead to performance loss or instability.

Empirical guidelines for practical deployment include:

• Start with nominal-model parameters (e.g., inertia, gain) estimated within $P_n(z)$3 of actual values.
• Select Q-filter cut-off ($P_n(z)$4 or digital $P_n(z)$5) just above maximum disturbance frequency but well below noise-dominant or resonance frequencies.
• For multi-axis machines, introduce modest non-diagonal nominal parameters (e.g., inertia cross-terms) to decouple and optimize disturbance rejection in specific channels without global noise penalty (Sariyildiz et al., 2019).
• Validate both regulation and trajectory-tracking mode to ensure the predicted boundedness and convergence guarantees.

Across applications and platforms, the parallel DOB emerges as a robust, theoretically grounded, and practical solution for broadband disturbance rejection, underpinned by explicit tuning rules and supported by rigorous stability proofs and extensive experimental validation (Sariyildiz et al., 2019, Sariyildiz et al., 2019, Sariyildiz, 2021, Kwon et al., 2022, Seo et al., 25 May 2026, Kayacan et al., 2021).