(106d) Experimental Demonstration of Closed-Loop Optimization of a Single Column Pressure Swing Adsorption (PSA)-Based Oxygen Concentrator Using Machine Learning

With COVID-19 still active in most parts of the world, the need for affordable oxygen concentrators for affected patients is on the rise, especially in developing countries. Beyond COVID-19, oxygen concentrators are critical in the rehabilitation of patients afflicted by a range of pulmonary disorders, collected termed Chronic Obstructive Pulmonary Disorder (COPD). Typically, PSA oxygen concentrators are medical devices that separate oxygen from atmospheric air using a cyclically operated Pressure Swing Adsorption (PSA) process carried out with beds of nitrogen selective zeolite beads. The current industry standard is to use a double column PSA design, where each column will take turns undergoing pressurization/adsorption and depressurization/desorption. Although the method is already efficient, the viability of a single-column design and a product storage tank has been researched for its reduced cost and size. Such a single-column design employs a variation of the Skarstorm cycle which involves four steps applied to feed air passing through a zeolite column: (a) column pressurization; (b) nitrogen adsorption at high pressure resulting in oxygen production; (c) column depressurization resulting in nitrogen desorption; and (d) column purge using a portion of the product oxygen from the product tank to further remove residual nitrogen from the zeolite column. A functional prototype of the unit described above has been successfully built and semi-optimized to test if the unit is able to produce oxygen at a purity above 90% at a flow of 5 slpm.

The operation of this cyclic process involves use of solenoid valves controlled by an Arduino Uno board with python code, which opens and closes valves at the inlet and exit of the single column and the product storage tank to carry out the four-step Skarstorm cycle. Traditional approaches to optimize oxygen concentration in PSA setups involve an extensive process of user trial and error. Further, these approaches are challenging to alter should changes in the dynamics of the PSA system cause a degradation of performance. To remedy these challenges, we implemented a recently developed nonlinear online MPC training algorithm with a feedforward neural network (FF-MPC) to optimize the four cycle step times (pressurization, adsorption, desorption, and purge). FF-MPC uses a generic artificial feedforward neural network with pre-specified hyper-parameters and trains the neural network in a closed-loop MPC framework by using the measured output (oxygen purity) obtained in response to the applied control actions (four step times). Consequentially, the neural network predictive performance is improved in real-time by applying online weight updates using the system’s state feedback.

Our preliminary results show that the FF-MPC algorithm successfully integrates with the Arduino code, indicated by closed-loop control of the solenoid valves. The closed-loop hardware

implementation consists of the following four steps: (1) the Arduino board receives a measurement of oxygen purity, (2) the incoming measurements are used to improve predictions of the FF model, (3) the FF-MPC algorithm optimizes a new set of cycle step time and updates the Arduino code, and finally (4) a new measurement of oxygen purity detected by the oxygen analyzer is sent to the Arduino code.

The resulting optimization process is favorable compared to heuristic tuning because the FF-MPC approach provides an optimization framework capable of learning the nonlinear relationship between cycle step times and oxygen concentration. Our experimental demonstration has the potential to enable easier tuning and operation of existing commercial oxygen concentrators.