Josip Zidar’s research while affiliated with University of Osijek and other places

Calibration of low-cost humidity sensors such as the HTS221TR is critical for accurate measurements, especially in smart devices. This study compares two calibration methods: machine learning (PyTorchNeural Network regression model) and optimization algorithm with Engineering Equation Solver. The critical role of temperature in humidity measurement emphasizes that it must be included for a valid calibration. The machine learning approach significantly reduced the average deviation of humidity, reaching ±2,5% compared to the original ±13,4%. Additionally, it aligned mean values along the identity line. However, the performance of the model varied across the different humidity ranges. Applying the model to real-world scenarios showed that the model underestimates humidity, likely due to the sensor's inherent tendency to overestimate humidity, especially at higher temperatures. Despite these challenges, both calibration methods offer simple and effective approaches for correcting low-cost sensor measurements, with machine learning enabling faster processing. This study not only improves the accuracy of the HTS221TR sensor, but also paves the way for more accurate and affordable humidity measurement technologies in general.

Dynamic voltage and frequency scaling (DVFS) is a technique used to optimize energy consumption in ultra-low-power embedded systems. To ensure sufficient computational capacity, the system must scale up its performance settings. The objective is to conserve energy in times of reduced computational demand and/or when battery power is used. Fast Fourier Transform (FFT), Cyclic Redundancy Check 32 (CRC32), Secure Hash Algorithm 256 (SHA256), and Message-Digest Algorithm 5 (MD5) are focused functions that demand computational power to achieve energy-efficient performance. Selected operations are analyzed from the energy consumption perspective. In this manner, the energy required to perform a specific function is observed, thereby mitigating the influence of the instruction set or system architecture. For stable operating voltage scaling, an exponential model for voltage calculation is presented. Statistical significance tests are conducted to validate and support the findings. Results show that the proposed optimization technique reduces energy consumption for ultra-low-power applications from 27.74% to up to 47.74%.

This paper presents a shock detection device for packages in the supply chain. The primary purpose is to identify package damage during storage, delivery, and handling. Additionally, products are likely to be damaged if dropped from a certain height, which sometimes does not appear on the package. By continuously measuring package vibrations and detecting shocks in the supply chain, consumers can gain an insight into the state of the product upon delivery. This paper presents the Smart Sticker implementation for ultra-low-power shock detection in the supply chain. The overall energy consumption must be kept as low as possible while continuously sensing the presence of shock to ensure that the Smart Sticker’s battery lasts as long as possible. The Smart Sticker functions in three modes to meet the established constraints: low-power, active, and data transfer mode. While detecting the shock, the low-power mode uses the least amount of energy needed. If the shock exceeds the threshold, the Smart Sticker enters active mode, stores the detected g force value in memory, and then switches back to low-power mode. Finally, employing Near Field Communication (NFC) and energy harvesting, the data transfer mode allows the consumer to read the recorded data. The results show that the Smart Sticker for shock detection performs according to set requirements and successfully monitors and detects shock for packages in the supply chain.

In the modern world different products and goods are available throughout the world thanks to the complex supply chain system. Often products are transported on long journeys with different transportation systems where products can be damaged or spoiled. Smart Sticker is a concept for product environmental condition monitoring that can resolve some problems in the supply chain. Smart Sticker will record product environmental data in the supply chain and enable producer/consumer product monitoring. Because of ultra-low power design, Smart Sticker component selection must satisfy ultra-low power specifications, besides standard accuracy, and real-time implementation. In this paper we give an overview of the necessary measured environmental parameters and the selection of sensors with emphasis on low power design. We provided a model for the calculation of the maximum operating time, which is applied for the two Smart Sticker instances with significantly different energy consumptions. In the worst-case scenario operation time is 198 days which can be increased with a higher capacity battery