Materials synthesis remains a critical bottleneck in developing innovations for energy storage, catalysis, electronics, and biomedical devices. Current synthesis design relies heavily on empirical trial-and-error methods guided by expert intuition, limiting the pace of materials discovery. To address this challenge, we present AlchemyBench, a comprehensive benchmark built upon a curated dataset of 17,667 expert-verified synthesis recipes from open-access literature.

State space models (SSMs) have emerged as an efficient alternative to Transformer models for language modeling, offering linear computational complexity and constant memory usage as context length increases. However, despite their efficiency in handling long contexts, recent studies have shown that SSMs, such as Mamba models, generally underperform compared to Transformers in long-context understanding tasks. To address this significant shortfall and achieve both efficient and accurate long-context understanding, we propose LongMamba, a training-free technique that significantly enhances the long-context capabilities of Mamba models. LongMamba builds on our discovery that the hidden channels in Mamba can be categorized into local and global channels based on their receptive field lengths, with global channels primarily responsible for long-context capability. These global channels can become the key bottleneck as the input context lengthens. Specifically, when input lengths largely exceed the training sequence length, global channels exhibit limitations in adaptively extend their receptive fields, leading to Mamba’s poor long-context performance. The key idea of LongMamba is to mitigate the hidden state memory decay in these global channels by preventing the accumulation of unimportant tokens in their memory. This is achieved by first identifying critical tokens in the global channels and then applying token filtering to accumulate only those critical tokens. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for Mamba’s long-context performance, significantly extending its operational range without requiring additional training. Our code is available at https://github.com/GATECH-EIC/LongMamba.

Effectively modeling relationships between distant elements in a long input sequence is an important task that remains challenging to this day. The state-of-the-art models for processing sequential data are self-attention-based transformer models. However, the computational complexity of self-attention is quadratic to the input sequence length, which often becomes the limiting factor in scaling the sequence length. Recently, state space model (SSM)-based global convolution models, which replace attention with convolution, have been found to be effective for modeling long sequences, with a sub-quadratic complexity using Fast Fourier Transform (FFT). However, they show sub-optimal performance on data-parallel accelerators like GPU, due to the regions of extremely low compute utilization with memory bandwidth-bound operations. To address this inefficiency, this paper proposes the Vandermonde matrix Generating Accelerator (VGA), a custom accelerator that performs FFT-based convolution in an area/power-efficient manner. VGA introduces Complex number Compute Units (CCUs) to fully utilize the high on-chip SRAM bandwidth, and parameters are generated on the fly to drastically reduce the required SRAM capacity. VGA achieves 76×(48×) higher area (power) efficiency than NVIDIA A100 GPU when executing the global convolution operator of H3, a state-of-the-art SSM-based model.

3D Gaussian splatting has recently emerged as a promising technique for novel view synthesis from sparse image sets, yet comes at the cost of requiring millions of 3D Gaussian primitives to reconstruct each 3D scene. This largely limits its application to resource-constrained devices and applications. Despite advances in Gaussian pruning techniques that aim to remove individual 3D Gaussian primitives, the significant reduction in primitives often fails to translate into commensurate increases in rendering speed, impeding efficiency and practical deployment. We identify that this discrepancy arises due to the overlooked impact of fragment count per Gaussian (i.e., the number of pixels each Gaussian is projected onto). To bridge this gap and meet the growing demands for efficient on-device 3D Gaussian rendering, we propose fragment pruning, an orthogonal enhancement to existing pruning methods that can significantly accelerate rendering by selectively pruning fragments within each Gaussian. Our pruning framework dynamically optimizes the pruning threshold for each Gaussian, markedly improving rendering speed and quality. Extensive experiments in both static and dynamic scenes validate the effectiveness of our approach. For instance, by integrating our fragment pruning technique with state-of-the-art Gaussian pruning methods, we achieve up to a 1.71$\times$ speedup on an edge GPU device, the Jetson Orin NX, and enhance rendering quality by an average of 0.16 PSNR on the Tanks\&Temples dataset. Our code is available at https://github.com/GATECH-EIC/Fragment-Pruning.

Model compression is widely adopted for edge inference of neural networks (NNs) to minimize both costly DRAM accesses and memory footprints. Recently, XOR-based model compression has demonstrated promising results to maximize compression ratio and minimize accuracy drop. However, XOR-based decompression alone produces bit errors and requires auxiliary data for error correction. To minimize model size and hence DRAM traffic, we propose an enhanced decompression algorithm and a low-cost hardware accelerator for it. Since not all errors are equal, our algorithm selects only important errors to correct with no accuracy drop. Compared with the baseline XOR compression scheme correcting all errors, the compressed model size of ResNet-18 and VGG-16 is reduced by 23% and 27% respectively. We also present a low-cost hardware implementation of on-line XOR decompression and error-correction logic built on Gemmini, an open-source systolic array accelerator, at the cost of only a 0.39% and 0.46% increase in area and power.

Predicting the remaining useful life (RUL) of systems is of importance in prognostics and health management to improve the maintenance capability in modern industries. Several approaches for RUL prediction using multiple sensory data assume the system's degradation in a specific form, which may not be applicable to the diversity of practical engineering systems. In this paper, a two-phase data driven framework is proposed. The first phase is an offline process of constructing the health index (HI) library from run-to-failure data using fuzzy similarity based data fusion, which enables to create the HIs without any assumptions. In the second phase, the HI of online is mapped onto the library created offline to find the best matching position using time series similarity search. This simultaneously enables the current condition of the system being informed as well as estimating its time left. The degradation simulation dataset of turbofan engine from NASA Prognostics Center is used to evaluate the potential application of the proposed framework.

This paper presents a new framework to improve processing time and accuracy for remaining useful life (RUL) prediction of a degrading equipment. Most of the existing machine learning-based RUL prediction approaches are often not robust to noise and outliers, requiring a large overhead processing to select best features and a complex procedure to build an accurate model. This paper explores using signal decomposition methods to eliminate the noise, and multiple deep learning techniques to minimize the outliers for a fast and accurate RUL prediction system. Specifically, we leverage either the wavelet transform (WT) or empirical mode decomposition (EMD) to extract time-frequency domain features, then build an ensemble deep neural network model that jointly utilize three different recurrent neural networks (RNNs) to predict the RUL of a degrading equipment. We evaluate the our proposed approach using aircraft gas turbine engine data from NASA, to show the effects of WT and EMD on the RUL prediction performance in terms of processing time and RUL accuracy. Experimental results demonstrate that our proposed approach significantly outperforms existing approaches.

In advanced manufacturing, tool condition monitoring is essential to improve the productivity of the manufacturing process. However, single model prediction is not satisfactory due to the complicated machining process that resulted in varying degradation of the tools. A multi-stage tool condition monitoring framework using gated recurrent unit network (Multi-GRU) is proposed in this paper to address the issue. A G RU classifier is first used to estimate the tool condition stage. Then different models are trained based on each stage to estimate the tool condition. A gun-drilling experiment under different cutting conditions is carried out to verify the efficiency of the proposed framework. The Multi-GRU framework is tested using extracted features for both tool state classification and tool wear estimation tasks. The experimental result shows that the Multi-GRU framework outperforms conventional single GRU model in the terms of overall accuracy in both classification and regression tasks. Besides, the experimental result also demonstrated that, with large enough classes, the classification results can approximate the tool wear with good accuracy.

Online machine learning algorithm has attracted increasing attention especially in the era of Industry 4.0. The reason is that traditional batch learning algorithm cannot deal with the streaming data produced by sensorized machines and make real-time decisions. In this paper, we propose a Feature-aware online learning approach of Support Vector Machine (FSVM) for classification problem. Usually, online learning algorithm has limited access to streaming data due to resource and computation constraints. In FSVM, we introduced a feature vector selection method to reduce the size of training dataset without losing key information and maintain an acceptable classification accuracy. Here, such small set of selected feature vectors is able to represent the original dataset. What is more, we can detect feature drifting by checking whether or not a new input data can be represented by the current feature vectors. We evaluate the performance of FSVM based on several realworld datasets. The results show that even train the SVM model with around 10% data, an acceptable misclassification rate can be reached.

An assumption on deterministic travel times of the vehicles might not be realistic for the dispatching problem of multiple Autonomous Intelligent Vehicles (AIVs). Thus, this paper presents a more accurate approach to solve this problem by considering stochastic travel times of the vehicles. The problem is modeled as a stochastic combinatorial optimization (SCO) problem that aims to minimize the total travel time and balance the utilization time of AIVs. Genetic Algorithm (GA) is then used to provide near-optimal solutions for the SCO. The obtained near-optimal results are then compared with the ones using conventional ruled-based approaches. The results showed that the GA solver could outperform the conventional ruled-based approaches in both deterministic and stochastic testing scenarios.

Imbalanced datasets are often encountered in process monitoring, where the data reflecting abnormal events like machine failures is less than the data reflecting normal events. The former is called the minority class and the later is referred as the majority class. Classical machine learning algorithms are still facing challenges in solving this problem. In order to improve the classification accuracy, oversampling techniques rebalance the dataset by supplying the minority class with synthetic samples. However, the latent sample spaces of both classes are broad, the majority class might be under-represented as well. In this paper, we propose a dual oversampling strategy (DOSS) to generate samples for both classes. For the majority class, synthetic samples are generated according to the data distribution, which is approximated by conditional Generative Adversarial Network (cGAN). For the minority class, Synthetic Minority Over-sampling Technique (SMOTE) is applied as the oversampling method. The proposed strategy is compared with others that either only the minority class is oversampled or both classes are oversampled with different strategies. Recall, G-mean and F-measure are used as the metrics. The experimental results on 12 benchmark datasets show the improved performance of our proposed strategy. DOSS is further applied to detect the faulty stages of an injection moulding machine where the prediction of DOSS achieves a better accuracy.

Tool condition monitoring (TCM) is a prerequisite to ensure high finishing quality of workpiece in manufacturing automation. One of the most important components in TCM system is tool wear estimation. How to achieve estimation with high accuracy is still an open question. In the past few decades, recurrent neural network (RNN) has shown a great success in learning long-term dependence of the sequential data. However, traditional RNNs (e.g., vanilla RNN, etc.) suffer gradient vanishing or exploding problem as well as long computational training time when the model is trained through back propagation through time (BPTT). To address these issues, we propose a gated recurrent units (GRU) based neural network to estimate the tool wear for tool condition monitoring. The GRU neural network can analyze time-series data on multiple time scales and can avoid gradient vanishing during training. A real-world gun drilling experimental dataset is used as a case study for tool condition monitoring in this paper. The performance of the proposed GRU based TCM approach is compared with other well-known models including support vector regression (SVR) and multi-layer perceptron (MLP). The experimental results show that the proposed GRU based TCM approach outperforms other competing models on this real-world gun drilling dataset.

An abbreviation is often used in scientific communications when a concept is needed to be addressed repeatedly. Although some concepts can be described with a single word, e.g., mechanics, phrases are commonly required for more elaborated concepts, e.g., quantum mechanics. It is believed that by extracting abbreviations from full texts on a scientific topic could help expand the domain knowledge of a novice. A text mining algorithm is used to extract the abbreviations from 247 full texts on predictive maintenance - a manufacturing topic - hosted on ScienceDirect. It is found that the abbreviations extracted are useful for generating domain-specific dendrograms, which can provide valuable information during literature review on a scientific area. This paper is the second of a multi-part study, namely, performing literature review using text mining.

Co-localization of mobile users combines methods of detecting nearby users and providing them interesting and useful services or information. By exploiting the massive use of smartphones, nearby users can be co-localized using only their captured ambient radio signals. In this paper, we propose a real-time co-localization system, in a centralized manner, that leverages co-located users with high accuracy. We exploit the similarity of radio frequency measurements from users' mobile terminal. We do not require any further information about them. Our co-localization system is based on a nonparametric Bayesian method called infinite Gaussian mixture model that allows the model parameters to change with observed input data. In addition, we propose a modified version of Gibbs sampling technique with an average similarity threshold to better fit user's group. We design our system in a completely centralized manner. Hence, it enables the network to control and manage the formation of the users' groups. We first evaluate the performance of our proposal numerically. Then, we carry out an extensive experiment to demonstrate the feasibility, and the efficiency of our approach with data sets from a real-world setting. Results on experiment favor our algorithm over the state-of-the-art community detection-based clustering method.

Nowadays people are carrying their mobile devices wherever they go, and as social beings they interact with others all day long. Thus, by exploiting this massive use of smart devices they provide a way to be co-located using only their captured environmental radio signals. In this paper, we design a co-location system that finds groups of people, in real-time, with high accuracy, by exploiting the similarity of their measured radio signals. Our method is based on a nonparametric Bayesian (NPB) method called infinite Gaussian mixture model (IGMM) that allows the model parameters to change with observed input data. This system is designed in a completely centralised manner. Hence, it enables the network to control and manage the formation of the all users' groups. We analyze the performance of our framework, in terms of clustering accuracy, with datasets from a real-world setting to demonstrate its feasibility. We also compare its performance against community detection based clustering method. Results on experiment with real datasets show a better accuracy favoring our approach against its counterpart.

Proximity information of users can be used in various applications (e.g., user interactions in social networks). Applying conventional works to room-level proximity detection is difficult because there is a limitation of the proximity range within 5 m. In this paper, we propose a room-level proximity detection method based on the similarity of received information of Wi-Fi signals between users. We use Wi-Fi signals in both 2.4 GHz band and 5 GHz band, and use relevant features that indicate the similarity of received signal strength (RSS) of beacon frames and access points (APs) sets from which users receives them. Through extensive experiments, in which we recognize whether or not users exist in the same room with a size approximately from 10 to 15 m square, we demonstrate that our proposed method can realize room-level proximity detection with high robustness to relative location of users and APs.

Nowadays, the population of elderly persons who live alone is increasing in Japan. Some systems are available to monitor those persons' activities, however, they have some limitations; physical burden and need to wear sensors. In our previous work, we have proposed a fall detection method using an array antenna. Array antenna is an antenna, which detects radio wave propagation by observing the direction of arrival (DOA) of the received signal. In the conventional method, it can only detect whether the activity is fall or not, so it is impossible to recognize which kind of non-fall activity the person does. In addition, the detection accuracy of the non-fall activity is not so high, because of the difficulty to detect the difference of fall and sit. In this paper, we propose an activity recognition system by using array antenna. We detect the person's activity (fall, walk, sit, and stand still). To improve the accuracy, we set thresholds for the continuous number of each activity, and correct the result which continues less than the threshold. Also, we use the state transition of the activity, and correct the impossible activity transition.

Now, aging society is a worldwide problem, and the population of people aged over 60 years is growing faster than any other age group. Therefore, monitoring services for elderly people are attracting a great deal of attention. We have proposed a fall detection method using a low resolution infrared array sensor to inform an unexpected falling in our previous work. However, knowing daily fundamental activities of elderly people is also important to prevent future accidents. In this paper, we propose an activity recognition method using a low resolution infrared array sensor. This sensor can detect temperature on a two dimensional area. From the viewpoint of general versatility (available in darkness), cost, size, privacy (low resolution), and availability (commercial off-the-shelf), this sensor is better than other sensing devices like video cameras, Doppler radars, acceleration sensors, and so on. In the proposed method, temperature distribution obtained from the sensor is analyzed and classified into five fundamental states: “No event”, “Stopping”, “Walking”, “Sitting”, and “Falling” (emergency situation). As a result of experiments, our proposed method achieved recognition accuracy of 100 %, 94.8 %, 99.9 %, and 78.6 % respectively. In particular, 100 % accuracy of “Falling” recognition was achieved.

Device-free passive (DFP) localization systems are a key solution for location-based services because they do not require any wireless device on a human body. Most of the existing DFP localization systems are based on the received signal strength (RSS) measurement only. However, the localization accuracy of RSS only-based systems is easily affected by the spatial and temporal variances of RSS due to multipath fading and noise, even in a static environment. In this paper, we propose a novel localization system for DFP using signal subspace eigenvectors from an antenna array. We present a fingerprinting technique using multiclass support vector machines (SVMs) based on a combination of array signal features with spatial and temporal averaging. We then evaluate the localization accuracy of our proposed system in different propagation environments: line-of-sight (LOS) and non-line-of-sight (NLOS). In addition, we analyze two types of receive antenna placement: centralized and distributed antennas. The experimental results show that the localization accuracy can be improved by the proposed system, particularly in the centralized antenna case. Moreover, they show that the proposed system can improve localization accuracy compared to the conventional RSS-only based system.

Nowadays, the population of the elderly people is increasing in Japan. The system to protect an elderly person who is living alone is needful. The bigger part of the accident in their house is falling. There are several kinds of products to detect a person's fall. However, there are some problems that it does not have enough detection range and privacy. Array antenna is an antenna, which detects radio wave propagation by observing the direction of arrival (DOA) of the signal. We developed previously the fall detection using array antenna. In the conventional fall detection method, it is needed to learn and observe the whole activity scenario, which includes how the person moves before and after the fall. As a result, when the unexpected fall scenario happens, it is not able to detect the fall correctly. In this paper, we propose a detection algorithm for unexpected fall using array antenna. We detect fall for every fixed time, and by the results of the detection, we decide whether the activity is falling or not. By using this method, it can detect the unexpected fall scenario, which is not learned. In addition, we use new features for support vector machine (SVM) to distinguish confusing activities and improve the fall detection accuracy.

Nowadays, aging society is a big problem and demand for monitoring systems is becoming higher. Under this circumstance, a fall is a main factor of accidents at home. From this point of view, we need to detect falls expeditiously and correctly. However, usual methods like using a video camera or a wearable device have some issues in privacy and convenience. In this paper, we propose a system of fall detection using a low resolution infrared array sensor. The proposed system uses this sensor with advantages of privacy protection (low resolution), low cost (cheap sensor), and convenience (small device). We propose four features and based on them, classify activities as either a fall or a non-fall using k-nearest neighbor (k-NN) algorithm. We show a proof-of-concept of our proposed system using a commercial-off-the-shelf (COTS) hardware. Results of experiments show the detection rate of higher than 94% irrespective of training data contains object's data or not.

Doppler radar-based fall detection has been attractive due to its low cost and high detection performance. However, fall detection based on Doppler signatures is affected by the spatial variance due to the multipath and non-line-of-sight (NLOS) effects, which has been one of the key issues for detection performance in current Doppler radar-based systems. Moreover, the drawbacks of Doppler radar are the limited measurement range and sensitivity to the target's falling directions. In this paper, we present a cooperative fall detection system that uses a Doppler radar and an array sensor which can be used even for multipath and NLOS environments. We analyze the impact of Doppler signatures in multipath and NLOS environments and account for undesirable Doppler measurements. We use the temporal-spatial characteristics of the signals using an array sensor and propose a novel fall detection system, which cooperates with the Doppler radar to enhance fall detection performance in multipath and NLOS environments. We evaluate the proposed system performance in a typical laboratory environment in LOS and NLOS conditions. The experimental results show that the proposed system reduces the false alarm rate and improves true positive rate. Moreover, our proposed system can significantly enhance the fall detection accuracy compared with the corresponding only Doppler radar-based approach.

We present a hidden Markov model based localization using array sensor. In this method, we use the eigenvector spanning signal subspace as a feature for location. The eigenvector does not depend on received signal strength (RSS) but on direction of arrival (DOA) of incident signals. As a result, the eigenvector is robust to fading and noise. In addition, the eigenvector is unique to the environment of propagation due to indoor reflection and diffraction of the electric wave. The conventional method based on fingerprinting does not take previous information into account. In this paper, we propose an algorithm that applies HMM to conventional fingerprinting of the eigenvector. This algorithm takes previous state of estimation into account by comparing the eigenvector obtained during observation with the one stored in the database. The database has the eigenvector obtained at each reference location according to setting in advance. In an indoor environment represented in a quantized grid, we decide the HMM transition probabilities denoting the possible moving range from previous estimation location. The most likely trajectory is calculated by means of the Viterbi algorithm. The results show that the localization accuracy is improved owing to the use of a possible moving range from the previous location.

In this paper we focus on improving state classification methods that can be implemented in elderly care monitoring systems. The authors group has previously proposed an indoor monitoring and security system (array sensor) that uses only one array antenna as the receiver. The clear advantages over conventional systems are improvement of privacy concern from the usage of closed-circuit television (CCTV) cameras, and elimination of installation difficulties. Our approach is different from the previous detection method which uses an array of sensors and a threshold that can classify only two states: nothing and something happening. In this paper, we present a state classification method that uses only one feature obtained from the radio wave propagation, and assisted by multiclass support vector machines (SVM) to classify the occurring states. The feature is the first eigenvector that spans the signal subspace of interest. The proposed method can be applied to not only indoor environments but also outdoor environments such as vehicle monitoring system. We performed experiments to classify seven states in an indoor setting: “No event,” “Walking,” “Entering into a bathtub,” “Standing while showering,” “Sitting while showering,” “Falling down,” and “Passing out;” and two states in an outdoor setting: “Normal state” and “Abnormal state.” The experimental results show that we can achieve 96.5 % and 100 % classification accuracy for indoor and outdoor settings, respectively.