Abstract - In modern industry, the integration of sensors with advanced artificial intelligence (AI) algorithms is essential for enhancing workflow efficiency and decision-making capabilities. This work introduces two innovative approaches that use these technologies to monitor and analyze liquid properties in real-time, shown in Figure 1. In the first approach (Figure 1 (a)), we designed a cuboid-shaped fluidic cell [1], fabricated from various materials using novel 3D printing techniques, featuring a vibrating membrane at its base with two external piezoelectric actuators attached. The second approach (Figure 1 (b)) employed micromachined plates driven by AlN piezoelectric films, fabricated using MEMS technology, meticulously designed to achieve quasi-free-free vibration in the (0,5) mode [2]. Both types of sensors used a two-port structure, one for actuation and the other for detection. The integration of these devices with AI techniques allowed us to use frequency responses in a range with multiple resonances, highly sensitive to fluid properties [3], while eliminating the need for complex electronics to process and acquire data.
Figure 1. (a) Top and cross-sectional views of the 3D-printed liquid cell with piezoelectric actuators. (b) Portable, low-cost viscometer-densimeter that included a MEMS microresonator, a microcontroller unit, and conditioning electronic circuits. The system incorporated a 3D-printed fluidic cell for injecting liquid into the sensor.
The spectra obtained with the sensors were subjected to various advanced data processing and machine learning techniques, performing an exhaustive search for the optimal combinations of hyperparameters that best fit the sensor data. Convolutional neural networks (CNNs) were found to be highly effective in working with frequency characteristics and estimating the viscosity and density of different types of liquids. In the second approach, these models were implemented on a microcontroller board, which also managed all electronics and communication with the sensor, resulting in a precise, compact, portable, and low-cost device.
Cell-based systems proved effective for monitoring the properties of aqueous solutions, achieving calibration errors below 2% and resolutions of 7.79 · 10-3 mPa·s for viscosity, and 1.09 · 10-3 g/mL for density. The microelectromechanical resonator-based instrument was capable to detect very small adulterations in olive oil with other vegetable oils, as low as 2%, with calibration and resolution errors of 0.47% and 0.14 mPa·s for viscosity, and 0.0331% and 9.25 · 10-5 g/mL for density. The calibration and resolution accuracies obtained were comparable to or exceeded those in the state of the art, and were on par with other commercial laboratory instruments of greater complexity, cost, and stationary nature.
Our findings demonstrate the significant potential of integrating sensors with machine learning techniques to achieve accurate detection of physical properties in fluids and address complex and critical industrial challenges, such as olive oil fraud. These advancements pave the way for the development of next-generation sensors that are not only accurate and reliable but also scalable and adaptable to diverse applications, providing valuable tools for the new industry.
