Medical Oxygen Concentrators (MOCs) are commericial devices which produce 90% oxygen (O
2) from ambient air using Rapid Pressure Swing Adsorption (RPSA) technology. RPSA processes are cyclic, and use materials (most commonly zeolites) which selectively adsorb one species in a gas mixture. The MOC market is large, estimated to be more than $1.8 billion in 2017, with many manufacturers trying to reduce the size and weight of ever smaller MOC devices. We recently designed and built a novel, single bed MOC device which features a single adsorber column concentrically inserted into a product storage tank. The single bed design reduces the overall size and weight of the MOC device and does not require the synchronization of multiple adsorbent beds found in other MOC devices. Operating RPSA systems is highly complex due to fast (
<10 sec) total cycle times, flow reversals, non-steady dynamic behavior and other effects, which makes controlling MOC devices both challenging and necessary. As MOC devices become more portable, they are also subject to changing operating conditions which further complicate their design and operation. In our previous work, we demonstrated that the single-bed design poses a multivariable control problem, and we developed a Model Predictive Control (MPC) algorithm in simulation which controls the RPSA process by manipulating the cycle step durations. The MPC features a linear model identified from a data-driven sub-space identification procedure and PseudoRandom Binary Sequence (PRBS) input signals as well as a convex quadratic optimization program with linear inequality constraints. The advantage of using a data-driven model identification technique is that the same procedure used in simulation can be repeated on the prototype unit, thus eliminating potential modeling errors in the controller formulation.
In this work, we present the implementation of this multivariable MPC algorithm on a lab-scale MOC prototype using Raspberry Pi® hardware. Python code and custom circuitry is used to both operate the RPSA cycle and solve the MPC quadratic program in real-time on a cycle-to-cycle basis. The model identification procedure uses measured output data, carefully designed PRBS-type input signals and the sub-space identification algorithm n4sid to generate the linear model used in the MPC. Implementation challenges such as hardware requirements, sensor dynamics, measurement error and task synchronization and imposed safeguards are discussed in detail. The closed-loop MOC-MPC system is able to operate reliably at the desired performance by operating the RPSA cycle and rejecting process disturbances. Several disturbance rejection and set point tracking case are presented to evaluate the MPC performance.
