On-device artificial intelligence has attracted attention globally, and attempts to combine the internet of things and TinyML (machine learning) applications are increasing. Although most edge devices have limited resources, time and energy costs are important when running TinyML applications. In this paper, we propose a structure in which the part that preprocesses externally input data in the TinyML application is distributed to the hardware. These processes are performed using software in the microcontroller unit of an edge device. Furthermore, resistor–transistor logic, which perform not only windowing using the Hann function, but also acquire audio raw data, is added to the inter-integrated circuit sound module that collects audio data in the voice-recognition application. As a result of the experiment, the windowing function was excluded from the TinyML application of the embedded board. When the length of the hardware-implemented Hann window is 80 and the quantization degree is $2^{-5}$, the exclusion causes a decrease in the execution time of the front-end function and energy consumption by 8.06% and 3.27%, respectively.

As the application fields for digital twins have expanded, various studies have been conducted with the objective of optimizing the costs. Among these studies, research on low-power and low-performance embedded devices has been implemented at a low cost by replicating the performance of existing devices. In this study, we attempt to obtain similar particle count results in a single-sensing device replicated from the particle count results in a multi-sensing device without knowledge of the particle count acquisition algorithm of the multi-sensing device. Through filtering, we suppressed the noise and baseline movements of the raw data of the device. In addition, in the process of determining the multi-threshold for obtaining the particle counts, the existing complex particle count determination algorithm was simplified to make it possible to utilize the look-up table. The proposed simplified particle count calculation algorithm reduced the optimal multi-threshold search time by 87% on average and the root mean square error by 58.5% compared to existing method. In addition, it was confirmed that the distribution of particle count from optimal multi-thresholds has a similar shape to that from multi-sensing devices.

Smart manufacturing utilizes digital twins that are virtual forms of their production plants for analyzing and optimizing decisions. Digital twins have been mainly developed as discrete-event models (DEMs) to represent the detailed and stochastic dynamics of productions in the plants. The optimum decision is achieved after simulating the DEM-based digital twins under various what-if decision candidates; thus, simulation acceleration is crucial for rapid optimum determination for given problems. For the acceleration of discrete-event simulations, adaptive abstraction-level conversion approaches have been previously proposed to switch active models of each machine group between a set of DEM components and a corresponding lookup table-based mean-delay model during runtime. The switching is decided by detecting the machine group’s convergence into (or divergence from) a steady state. However, there is a tradeoff between speedup and accuracy loss in the adaptive abstraction convertible simulation (AACS), and inaccurate simulation can degrade the quality of the optimum (i.e., the distance between the calculated optimum and the actual optimum). In this article, we propose a simulation-based optimization that optimizes the problem based on a genetic algorithm while tuning specific hyperparameters (related to the tradeoff control) to maximize the speedup of AACS under a specified accuracy constraint. For each individual, the proposed method distributes the overall computing budget for multiple simulation runs (considering the digital twin’s probabilistic property) into the hyperparameter optimization (HPO) and fitness evaluation. We propose an efficient HPO method that manages multiple Gaussian process models (as speedup-estimation models) to acquire promising optimal hyperparameter candidates (that maximize the simulation speedups) with few attempts. The method also reduces each individual’s exploration overhead (as the population evolves) by estimating each hyperparameter’s expected speedup using previous exploration results of neighboring individuals without actual simulation executions. The proposed method was applied to optimize raw-material releases of a large-scale manufacturing system to prove the concept and evaluate the performance under various situations.

Typical training procedures involve read and write operations for weight updates during backpropagation. However, on-device training on microcontroller units (MCUs) presents two challenges. Firstly, the on-chip SRAM has insufficient capacity to store the weight. Secondly, the large flash memory, which has a constraint on write access, becomes necessary to accommodate the network for on-device training on MCUs. To tackle these memory constraints, we propose a partial weight update technique based on gradient delta computation. The weights are stored in flash memory, and a part of the weight to be updated is selectively copied to the SRAM from the flash memory. We implemented this approach for training a fully connected network on an on-device MNIST digit classification task using only 20KB SRAM and 192KB flash memory on an MCU. The proposed technique achieves reasonable accuracy with only 18.52% partial weight updates, which is comparable to state-of-the-art results. Furthermore, we achieved a reduction of up to 46.9% in the area-power-delay product compared to a commercially available high-performance MCU capable of embedding the entire model parameter, taking into account the area scale factor.