We describe a method for precisely regulating the gradient magnet power supply (GMPS) at the Fermilab Booster accelerator complex using a neural network trained via reinforcement learning. We demonstrate preliminary results by training a surrogate machine-learning model on real accelerator data to emulate the GMPS, and using this surrogate model in turn to train the neural network for its regulation task. We additionally show how the neural networks to be deployed for control purposes may be compiled to execute on field-programmable gate arrays (FPGAs), and show the first machine-learning based control algorithm implemented on an FPGA for controls at the Fermilab accelerator complex. As there are no surprise latencies on an FPGA, this capability is important for operational stability in complicated environments such as an accelerator facility.

The implementation of damage-detection methods for continuously assessing structural integrity entails systems with attractive features such as storage capabilities, memory capacity, computational complexity and time-consuming processing. In this sense, embedded hardware platforms are a promising technology for developing integrated solutions in Structural Health Monitoring. In this paper, design, test, and specifications for a standalone inspection prototype are presented, which take advantage of piezo-diagnostics principle, statistical processing via Principal Component Analysis (PCA) and embedded systems. The equipment corresponds to a piezoelectric active system with the capability to detect defects in structures, by using a PCA-based algorithm embedded in the Odroid-U3 ARM Linux platform. The operation of the equipment consists of applying, at one side of the structure, wide guided waves by means of piezoelectric devices operated in actuation mode and to record the wave response in another side of the structure by using the same kind of piezoelectric devices operated in sensor mode. Based on the nominal response of the guide wave (no damages), represented by means of a PCA statistical model, the system can detect damages between the actuated/sensed points through squared prediction error (Q-statistical index). The system performance was evaluated in a pipe test bench where two kinds of damages were studied: first, a mass is added to the pipe surface, and then leaks are provoked to the pipe structure by means of a drill tool. The experiments were conducted on two lab structures: (i) a meter carbon-steel pipe section and (ii) a pipe loop structure. The wave response was recorded between the instrumented points for two conditions: (i) The pipe in nominal conditions, where several repetitions will be applied to build the nominal statistical model and (ii) when damage is caused to the pipe (mass adding or leak). Damage conditions were graphically recognized through the Q-statistic chart. Thus, the feasibility to implement an automated real-time diagnostic system is demonstrated with minimum processing resources and hardware flexibility.

In this paper the results of using a Multi-agent system (MAS) for Structural Health Monitoring (SHM) are detailed. A study between different MAS architectures reported in literature is presented in order to select and adapt the most adequate one for SHM tasks. Requirements are established according to recent solutions, where main parameters are type and number of sensors and communication protocols, among others. MAS technique uses several intelligent agents, that are algorithms able to act in a reactive or active way. Their action depends on surrounding environment or collected data. These agents can work in a decentralized way, searching the fulfillment of an individual goal or they can work with another system to achieve a common goal. Decision is based on their internal state (beliefs, goals and commitments). MAS’ effectiveness depends on the interconnection between the agents. Type of agents is defined according to its communication method and protocol, common and individual goals, among others. Decentralization and versatility are two important characteristics of MAS technique useful to solve SHM problem. This is one of the main motivations to consider this technique to be a good approach for the studied problem. A benchmark numerical model, which consists of a metallic framework, was used to validate and demonstrate the feasibility of the selected architecture for SHM