Research
Closed-Loop Dynamic Validator Node Scaling in Private Substrate Blockchains Using Takagi-Sugeno Fuzzy Inference
Private blockchains typically run with a fixed number of validator nodes, but the workloads they serve (such as IoT sensor data) rise and fall over time. Too many validators on a light workload…

- arXiv
- 2607.07901
- Published
- 2026-07-08
- Authors
- Thandile Nododile, Ayinde M. Usman, Clement N. Nyirenda
AI summary
Overview
Research area: Blockchain infrastructure management — specifically adaptive validator-node provisioning in private (permissioned) Substrate chains, using fuzzy control theory (Takagi-Sugeno inference) rather than fixed thresholds.
Technical level: Advanced. The paper assumes familiarity with permissioned blockchain consensus (AURA/GRANDPA), fuzzy inference systems, and standard statistical testing; the prose is clear but the concepts are specialised.
Scope: A single paper presenting, calibrating, and empirically evaluating a 27-rule Takagi-Sugeno fuzzy controller that reads three live chain metrics and autonomously adds or removes validator nodes on a 10-node private Substrate testbed storing hashed smart water meter data.
What This Paper Is About
Private blockchains typically run with a fixed number of validator nodes, but the workloads they serve (such as IoT sensor data) rise and fall over time. Too many validators on a light workload wastes computing resources; too few on a heavy workload slows block production and delays finalisation.
The authors' goal is to replace crisp, fixed thresholds with a fuzzy controller that reads block production time, block size, and active node count, computes an efficiency score and an action value, and then actually activates or deactivates validators in a closed loop — aiming for stable scaling instead of the oscillatory behaviour they attribute to threshold-based rules.
Key Contributions
-
A closed-loop scaling system that acts, not just advises. The TS controller triggers real validator activation with peer-discovery verification and deactivation via SIGTERM, rather than only emitting a recommendation.
-
An empirical recalibration method for the fuzzy membership functions. Instead of anchoring linguistic terms to theoretical extremes, the authors anchor them to the observed operating range of the testbed (for example, moving the Block Time "High" peak from a predicted 20 s to the observed worst case of 14 s, and the Node Count "Many" peak from the chain ceiling of 10 to 8). A generalisable anchoring rule plus worked examples across different slot durations (6 s, 3 s, 12 s) are provided.
-
Two-directional closed-loop validation. Separate experiments demonstrate Scale Up under rising load and Scale Down under over-provisioning, both converging to the same six-validator equilibrium at the same mean block time of 8.4 s.
-
A comparative evaluation against three threshold baselines (Conservative at >12 s, Moderate at >10 s, Aggressive at >8 s), using Welch's t-tests and Cohen's d effect sizes across three operational regimes, plus decision-flip counts as a stability metric.
Main Findings
-
Configurations produce statistically distinct operating profiles. Across 5 repeats each of 4, 7, and 10 active validators, mean block times were 13.92 s, 7.78 s, and 6.00 s respectively, with within-run pooled standard deviations of 2.27 s, 1.80 s, and 0.00 s. Mean efficiency scores were 44.07%, 54.82%, and 50.00%; mean action values were 0.639, 0.435, and 0.500; mean blocks produced were 27.2, 48.6, and 69.4. One-way ANOVA gave F = 7229 (block time), F = 1042 (efficiency), and F = 2215 (action value), all with p < 0.001. The 4-vs-10 block-time Welch test produced t = 71.37, the largest magnitude reported.
-
Measured block times track the effective-slot formula. The nominal targets 6 × 10/n are 15.0 s, 8.57 s, and 6.0 s; measured means of 13.92 s, 7.78 s, and 6.00 s are consistent with these once adjacent filled slots partially offset gaps. The 10-node case is exactly constant at 6.00 s because all authority slots are filled.
-
The controller scaled up autonomously under growing load. In the 1200-second seven-phase unified experiment, Scale Up fired at t = 195 s (4 to 5 validators, peer discovery in 2.04 s, block time falling from 14.4 s to 10.2 s) and at t = 240 s (5 to 6 validators, block time falling from 13.0 s to 8.4 s) — reductions of 4.2 s and 4.6 s, landing within 1.8 s and 1.6 s of the 12.0 s and 10.0 s targets.
-
It then held a stable equilibrium. After t = 240 s the six-node configuration held for the remaining 960 s at 8.4 s block time. Recommendation distributions were 100% Maintain in Phases 1, 3, 4, 5, 6, and 7, and 83.3% Maintain / 16.7% Scale Up in Phase 2. The authors' hypotheses H1a, H1b, and H2 are reported as supported; H1c (Scale Down under this workload) was not triggered and was demonstrated in a separate experiment.
-
Scale Down was demonstrated separately from an over-provisioned start. Starting at eight active validators under light load, the action value held at 0.250 (Scale Down zone) and the controller deactivated a validator at t = 60 s (8 to 7) and again at t = 105 s (7 to 6). After convergence, block time averaged 8.4 s (oscillating 6.0–10.8 s) and action value averaged 0.46, inside the Maintain band. The distribution was 91.7% Scale Down during the idle phase and 100% Maintain thereafter.
-
Only a small fraction of the 27 rules fire in practice. Six rules reached firing strength ≥ 0.1 (R11, R14, R17, R20, R23, R26), with R12 adding a seventh in the over-provisioned run. R20 dominated the idle phase; R17 and R26 drove the heavy and maintain-high phases as block size grew; R11 took over when block size returned to baseline. Unfired rules correspond to states the testbed cannot produce, such as high block time combined with many active validators.
-
The stability advantage over thresholds is large. On the identical 1200-second workload, the TS controller produced 2 Scale Up, 0 Scale Down, and 4 decision flips, with a final block time of 8.4 s and final efficiency of 67.9%. Conservative produced 3 up, 1 down, 4 flips, 12.0 s, 40.0%. Moderate produced 18 up, 14 down, 29 flips, 7.9 s, 81.0%. Aggressive produced 17 up, 12 down, 24 flips, 11.0 s, 50.0%. Against Conservative, TS block times were significantly lower in all three regimes (d = −2.10 in Maintain, −2.11 in Scale Down, −0.58 in Scale Up, all p < 0.001). Against Moderate it was slightly slower in Scale Up (t = 2.56, p = 0.012, d = 0.39) with non-significant differences in Maintain (p = 0.077) and Scale Down (p = 0.068). Against Aggressive, Scale Up showed t = 2.56, p = 0.012, d = 0.39; Maintain t = −3.11, p = 0.003, d = −0.73; Scale Down p = 0.068.
-
The cooldown alone does not explain the stability. All controllers were held to the same 30-second stabilisation window, so the authors attribute the difference to fuzzy inference rather than to cooldown policy.
Methodology in Plain English
The team built a private Substrate blockchain with 10 validator authorities, a 6-second AURA slot, and GRANDPA finality. Sensor data files — daily CSVs of real smart water meter readings from the Queensland Government open data portal — are hashed with SHA-256 and only the 256-bit hash is written on-chain through a custom pallet; the raw files stay off-chain. A separate Python process queries the chain over JSON-RPC every 5 seconds to read block production time, block size, and how many validators are currently active.
Those three numbers are fed into a fuzzy controller. Each is described with three overlapping triangular curves (low/medium/high, small/medium/large, few/moderate/many) so that a value like 9.6 s belongs partly to "medium" and partly to "high" rather than triggering a hard switch. All 27 combinations of the three variables have a rule, and each rule carries two constants: an efficiency score and an action value. The controller multiplies the membership degrees (product t-norm) to get each rule's firing strength, then takes a weighted average of the consequents — a zero-order Takagi-Sugeno system, which the authors note completes in sub-millisecond time. An action value below 0.3 means scale down, 0.3 up to 0.7 means maintain, and 0.7 or above means scale up, bounded between 4 and 10 active validators. After any action, a 30-second window suppresses further decisions.
Crucially, the membership curves were not set from textbook extremes but from what the testbed actually did. The load generator was calibrated to 75 extrinsics per second as the reliably sustainable cap, and the seven-phase workload ramps from 1 ext/s to 75 ext/s and back down over 1200 seconds. Three fixed-threshold controllers were run on the identical workload, and block times were split into Scale Up, Maintain, and Scale Down regimes for Welch's t-tests with Cohen's d, complemented by counting how often each controller flipped between adjacent recommendations.
Why This Matters
Impact on research. Prior work from this group and others established blockchain-IPFS architectures for IoT data integrity but always with static validator sets. This paper argues, to the best of the authors' knowledge, that it is the first to apply fuzzy inference to closed-loop blockchain node orchestration, and it contributes a reusable calibration heuristic (anchor linguistic terms to observed operating range, not theoretical bounds) that lets the same 27 rules transfer to chains with different slot durations.
Real-world applications:
- Smart utility metering, where the workload follows daily and seasonal consumption patterns — the exact data source used here.
- Enterprise permissioned chains for supply chain or healthcare audit trails, where validator capacity is expensive and demand is bursty.
- Edge and IoT gateway deployments where a lightweight controller (sub-millisecond inference, 5-second polling) must run without heavy computation.
- Cost governance for cloud-hosted validator fleets, where over-provisioning is a direct recurring expense and churn itself costs time in keystore setup, peer discovery, and chain sync.
Industry relevance. The headline number for operators is stability: 4 decision flips versus 24 to 29. Each scaling action carries real operational cost, so an order-of-magnitude reduction in churn at comparable block times is a practical argument for fuzzy control over the threshold alarms that are standard in distributed-systems monitoring.
Future Directions
- Online adaptation of the membership functions. The authors propose an "evolving" variant that recalibrates the fuzzy parameters at runtime rather than through offline empirical anchoring, which would remove the manual calibration step that currently ties the controller to a specific deployment.
- Stronger baseline set. Comparison against hysteresis-based controllers and PID controllers is explicitly named as future work — hysteresis in particular is the obvious competitor for oscillation suppression and is not tested here.
- Extension to other infrastructure dimensions. Complementary fuzzy subsystems for storage configuration and archival management are proposed, broadening the controller from validator count to broader node provisioning.
- Robustness beyond one testbed. The rule base contains rules that never fire in this architecture, and the study covers a single 10-authority network with a 6-second slot and one data pallet; behaviour under different consensus mechanisms, adversarial conditions, or validator failures is not reported.
Target Audience
Researchers and engineers working on permissioned blockchain infrastructure, consortium or enterprise chain operations, and blockchain-IoT data pipelines — particularly those who own the operational cost of validator fleets. It also suits readers interested in applied fuzzy control outside classical mechanical systems, and graduate students looking for a worked example of closing the loop between a soft-computing controller and live distributed-system telemetry. Readers without background in consensus slot mechanics or fuzzy inference will find the system sections dense.
Authors’ abstract
Private blockchain networks run with fixed node configurations that cannot adapt to changing workload conditions. Too many nodes serving a light workload waste resources; too few nodes facing heavy demand slow block production and degrade finalisation. The right validator count is hard to determine, as it depends on overlapping factors that shift over time. This paper presents a Takagi-Sugeno (TS) fuzzy inference system that reads live blockchain parameters (block production…
The full abstract on arXiv