Self-Tuning Fully-Connected PID Neural Network System for Distributed Temperature Sensing and Control of Instrument with Multi-Modules

Zhen Zhang, Cheng Ma, Rong Zhu2016 · DOI

TL;DR

This paper introduces a novel self-tuning, fully-connected PID neural network (FCPIDNN) system for precise temperature control in complex, multi-module instruments. The system achieves global optimal temperature regulation without needing a system model and can dynamically reconfigure itself to handle actuator failures, enhancing reliability.

What Problem Does This Paper Address?

Modern instruments are increasingly miniaturized and modularized, leading to a high density of components within a single device. This creates a significant challenge for temperature control: the internal environment is complex, with multiple modules generating heat and having varying thermal characteristics. Maintaining a stable and uniform temperature across these modules is crucial for their reliable operation, especially when they require different temperature setpoints or when external conditions change. Traditional control methods often struggle with this "multi-input, multi-output" (MIMO) problem due to strong coupling between modules and time-varying internal structures, which makes accurate system modeling difficult. Existing approaches, like independent PID controllers for each module, fail to account for these interdependencies, leading to steady-state errors and suboptimal performance.

Methods

The core innovation is a Fully-Connected Proportional-Integral-Derivative Neural Network (FCPIDNN). This system builds upon the concept of PID Neural Networks (PIDNNs), which integrate PID control logic within a neural network structure, thus avoiding the need for explicit system modeling. The FCPIDNN enhances this by:

  • Fully-Connected Architecture: It connects the outputs of multiple individual PIDNN controllers (one for each module) to a final layer that drives the cooling fans. This allows for coordinated control across all modules, addressing the MIMO coupling issue.
  • Self-Tuning Mechanism: The network's weights are updated online using a back-propagation algorithm. This allows the system to continuously learn and adapt to changing conditions and optimize its performance in real-time, aiming for a global optimum.
  • Reconfiguration on Failure: A key feature is the system's ability to self-reconfigure when an actuator (cooling fan) fails. The FCPIDNN can adapt its control strategy to compensate for the loss of a fan, maintaining temperature control for the remaining modules.

Datasets

Not discussed in the paper. The experiments were conducted on a custom-built instrument mockup.

Key Insights and Ideas

  • MIMO Control Without Models: The FCPIDNN demonstrates that complex MIMO temperature control can be achieved effectively without relying on explicit mathematical models of the instrument's thermal dynamics. This is a significant advantage for systems with intricate and hard-to-model behaviors.
  • Global Optimization through Interconnection: By fully connecting the PIDNN controllers and the actuators, the system moves beyond independent control loops. This interconnectedness allows for a global optimization of temperature across all modules, mitigating the interference issues seen with separate controllers.
  • Robustness via Self-Reconfiguration: The ability to dynamically reconfigure the control strategy in response to actuator failure is a major step towards highly reliable and fault-tolerant systems. This "smart" adaptation ensures continued operation even under adverse conditions.
  • Convergence to Global Minima: The experimental results and the convergence analysis suggest that the cost function consistently decreases, indicating that the self-tuning process effectively drives the system towards optimal temperature control.

Key Results

The paper presents experimental results from a custom instrument mockup with six modules.

  • Improved Performance over PIDNN: Compared to using six independent PIDNN controllers, the FCPIDNN achieved simultaneous convergence of all module temperatures to their targets without steady-state errors. The independent PIDNNs showed significant steady-state errors due to mutual interference.
  • Dynamic Target Changes: The FCPIDNN successfully adapted to in-process changes in target temperatures for individual modules and groups of modules, demonstrating its self-tuning capability.
  • Actuator Failure Resilience: When one cooling fan failed, the FCPIDNN was able to modulate the temperatures of the remaining modules to their setpoints, showcasing its self-reconfiguration and robustness. For example, when fan 2 failed, temperatures were still modulated to the setting targets.

Limitations and Future Work

  • Complexity of Implementation: While effective, the FCPIDNN involves a neural network structure that might require significant computational resources for implementation in highly constrained embedded systems.
  • Generalizability: The experiments were conducted on a specific mockup. Further validation on diverse instrument designs and operating conditions would be beneficial.
  • Future Work: The authors suggest that the FCPIDNN system is applicable to a wide range of distributed measurement and control systems, including integration electronics and building environments, indicating potential for broader application. The paper also implies that further exploration of its self-reconfiguration capabilities under various failure modes could be a fruitful area of research.

Summary updated

BibTeX
@misc{zhang2016self,
  author = {Zhen Zhang and Cheng Ma and Rong Zhu},
  title = {{Self-Tuning Fully-Connected PID Neural Network System for Distributed Temperature Sensing and Control of Instrument with Multi-Modules}},
  year = {2016},
  doi = {10.3390/s16101709},
}
Figure