Commercial exoskeletons currently utilize multiple sensors, including inertial measurement units, electromyography sensors, and torque/force sensors, to detect human motion. While these sensors improve motion recognition by leveraging their unique strengths, they can also lead to discomfort due to direct skin contact, added weight, and complex wiring. In this paper, we propose a simplified motion recognition method that relies solely on encoders embedded in the motors. Our approach aims to accurately classify various movements by learning their distinctive features through a deep learning model. Specifically, we employ a convolutional neural network algorithm optimized for motion classification. Experimental results show that our model can effectively differentiate between movements such as standing, lifting, level walking, and inclined walking, achieving a test accuracy of 98.76%. Additionally, by implementing a sliding window maximum algorithm that tracks three consecutive classifications, we achieved a real-time motion recognition accuracy of 97.48% with a response time of 0.25 seconds. This approach provides a cost-effective and simplified solution for lower limb motion recognition, with potential applications in rehabilitation-focused exoskeletons.

CNN is one of the deep learning technologies useful for image-based pattern recognition and classification. For machining processes, this technique can be used to predict machining parameters and surface roughness. In electrical discharge machining (EDM), the machined surface is covered with many craters, the shape of which depends on the workpiece material and pulse parameters. In this study, CNN was applied to predict EDM parameters including capacitor, workpiece material, and surface roughness. After machining three metals (brass, stainless steel, and cemented carbide) with different discharge energies, images of machined surfaces were collected using a scanning electron microscope (SEM) and a digital microscope. Surface roughness of each surface was then measured. The CNN model was used to predict machining parameters and surface roughness.

In the semiconductor manufacturing industry, efficient operation of wafer transfer robots has a direct impact on productivity and product quality. Ball screw misalignment anomalies are a critical factor affecting precision transport of robots. Early diagnosis of these anomalies is essential to maintaining system efficiency. This study proposed a method to effectively diagnose ball screw misalignment anomalies using 1D-CNN and 2D-CNN models. This method mainly uses binary classification to distinguish between normal and abnormal states. Additionally, explainable artificial intelligence (XAI) technology was applied to interpret diagnostic decisions of the two deep learning models, allowing users to convince prediction results of the AI model. This study was based on data collected through acceleration sensors and torque sensors. It compared accuracies of 1D-CNN and 2D-CNN models. It presents a method to explain the model"s predictions through XAI. Experimental results showed that the proposed method could diagnose ball screw misalignment anomalies with high accuracy. This is expected to contribute to the establishment of reliable abnormality diagnosis and preventive maintenance strategies in industrial sites.

In this study, we proposed an AI-algorithm for face mask recognition based on the MobileNetV2 network to implement automatic door control in intensive care units. The proposed network was constructed using four bottleneck blocks, incorporating depth-wise separable convolution with channel expansion/projection to minimize computational costs. The performance of the proposed network was compared with other networks trained with an identical dataset. Our network demonstrated higher accuracy than other networks. It also had less trainable total parameters. Additionally, we employed the CVzone-based machine learning model to automatically detect face location. The neural network for mask recognition and the face detection model were integrated into a system for real-time door control using Arduino. Consequently, the proposed algorithm could automatically verify the wearing of masks upon entry to intensive care units, thereby preventing respiratory disease infections among patients and medical staff. The low computational cost and high accuracy of the proposed algorithm also provide excellent performance for real-time mask recognition in actual environments.

Predicting fall risk is necessary for rescue and accident prevention in the elderly. In this study, deep learning regression models were used to predict the acceleration sum vector magnitude (SVM) peak value, which represents the risk of a fall. Twenty healthy adults (aged 22.0±1.9 years, height 164.9±5.9 cm, weight 61.4±17.1 kg) provided data for 14 common daily life activities (ADL) and 11 falls using IMU (Inertial Measurement Unit) sensors (Movella Dot, Netherlands) at the S2. The input data includes information from 0.7 to 0.2 seconds before the acceleration SVM peak, encompassing 6-axis IMU data, as well as acceleration SVM and angular velocity SVM, resulting in a total of 8 feature vectors used to model training. Data augmentations were applied to solve data imbalances. The data was split into a 4 : 1 ratio for training and testing. The models were trained using Mean Squared Error (MSE) and Mean Absolute Error (MAE). The deep learning model utilized 1D-CNN and LSTM. The model with data augmentation exhibited lower error values in both MAE (1.19 g) and MSE (2.93g²). Low-height falls showed lower predicted acceleration peak values, while ADLs like jumping and sitting showed higher predicted values, indicating higher risks.

Recently, the estimation of joint kinetics such as joint force and moment using wearable inertial sensors has received great attention in biomechanics. Generally, the joint force and moment are calculated though inverse dynamics using segment kinematic data, ground reaction force, and moment. However, this approach has problems such as estimation error of kinematic data and soft tissue artifacts, which can lead to inaccuracy of joint forces and moments in inverse dynamics. This study aimed to apply a recurrent neural network (RNN) instead of inverse dynamics to joint force and moment estimation. The proposed RNN could receive signals from inertial sensors and force plate as input vector and output lower extremity joints forces and moments. As the proposed method does not depend on inverse dynamics, it is independent of the inaccuracy problem of the conventional method. Experimental results showed that the estimation performance of hip joint moment of the proposed RNN was improved by 66.4% compared to that of the inverse dynamics-based method.

Bell’s palsy is a disease that occurs primarily between ages of 15 and 60, especially in middle-aged individuals. Although this disease gradually recovers within weeks to months, recurrence and permanent sequelae are possible. Its causes are diverse and unclear. Appropriate treatment is unknown, threatening lives of patients with this condition. In this study, we measured the degree of facial paralysis in a model of Bell’s palsy patients using OpenCV and the H.B grade measurement method and classified measured values according to H.B grade classification. This enabled prediction of the type and risk of diseases that might occur depending on the degree of facial paralysis. Additionally, we utilized more coordinate data to confirm movement of facial muscles by region to address limitations of the Nottingham system measurement method. We graded the level of this movement to enable intuitive confirmation and confirmed differences between existing Nottingham system and the H.B grade. This simple system could determine the level of paralysis in patients with Bell’s palsy and their corresponding risk level for related diseases. It enables information on causative disease of patients with Bell’s palsy to be quickly obtained, enabling prompt treatment and support.

Falls are common among older people. Age-related changes in toe strength and force steadiness may increase fall risk. This study aimed to evaluate the performance of a fall risk prediction model using toe strength and force steadiness data as input variables. Participants were four healthy adults (25.5±1.7 yrs). To indirectly reproduce physical conditions of older adults, an experiment was conducted by adding conditions for weight and fatigue increase. The maximal strength (MVIC) was measured for 5 s using a custom toe dynamometer. For force steadiness, toe flexion was measured for 10 s according to the target line, which was 40% of the MVIC. A one-leg-standing test was performed for 10 s with eyes-opened using a force plate. Deep learning experiments were performed with seven conditions using long short-term memory (LSTM) algorithms. Results of the deep learning model were randomly mixed and expressed through a confusion matrix. Results showed potential of the model"s fall risk prediction with force steadiness data as input variables. However, experiments were conducted on young adults. Additional experiments should be conducted on older adults to evaluate the predictive model.

A clean room is used for adjusting the concentration of suspended particles using an air-conditioner. It has a fan-filter unit combining a centrifugal fan and a high-efficiency particulate air filter that purifies the outside air and directly affects its cleanliness. Defects in these systems are typically detected using special sensors for each fault, which can be costly. Therefore, this paper proposes a system for diagnosing defects in the fan-filter unit using a single differential sensor and deep learning. The fan-filter unit is part of the air-conditioning system, and it is usually defective in bearings, filters, and motors. These faults include ball wear, internal bearing contamination, filter contamination, and motor speed changes. Each defect was artificially induced in experiments, and the differential pressure data of each defect was learned using a long short-term memory (LSTM) deep learning algorithm. The results of deep learning experiments generated by randomly mixing data five times were presented using a confusion matrix, and the results showed an accuracy of 87.2±2.60%. Therefore, the possibility of diagnosing defects in the fan-filter unit using a single sensor was confirmed.

As the size of semiconductor devices gradually decreases, it is important to measure and analyze semiconductor devices, to improve the image quality of semiconductors. We use VDSR, one of the Super-Resolution methods to improve the quality of semiconductor devices’ SEM images. VDSR is also a convolutional neural network that can be optimized with various parameters. In this study, a VDSR model for semiconductor devices’ SEM images was optimized using parameters such as depth of layers and amount of training data. Meanwhile, the quantitative evaluation and the qualitative evaluation did not match at the low scale factor. To solve this problem, we proposed an MTF measurement method using the slanted edge for better quantitative evaluation. This method was verified by comparing the results with the PSNR and SSIM index results, which are known as quality indicators. Based on the results, it was confirmed that using the MTF value could be a better approach for the evaluation of SEM images of the semiconductor device than using PSNR and SSIM.

There have been frequent fatal accidents of firefighters at fire scenes. A firefighting robot can be an alternative to humans at a fire scene to reduce accidents. As a critical function of the firefighting robot, it is mandatory to autonomously detect a fire spot and shoot water. In this research, a deep learning model called YOLOv7 was employed based on thermal images to recognize the shape and temperature information of the fire. Based on the results of the test images, which were not used for learning purposes, a recognition rate of 99% was obtained. To track the recognized fire spot, a 2-DOF pan-tilt actuation system with cameras was developed. By using the developed system, a moving target can be tracked with an error of 5%, and a variable target tracking test by alternately covering two target braziers showed that it takes about 1.5 seconds to track changing targets. Through extinguishment experiments with a water spray mounted on the pan-tilt system, it was observed that the temperature of the brazier dropped from 600 degrees to 13 degrees. Based on the obtained data, the feasibility of a robotic firefighting system using image recognition was confirmed.

Recently, in-depth studies on sensors of autonomous vehicles have been conducted. In particular, the trend to pursue only camera-based autonomous driving is progressing. Studies on object detection using IR (Infrared) cameras is essential in overcoming the limitations of the VIS (Visible) camera environment. Deep learning-based object detection technology requires sufficient data, and data augmentation can make the object detection network more robust and improve performance. In this paper, a method to increase the performance of object detection by generating and learning a high-resolution image of an infrared dataset, based on a data augmentation method based on a Generative Adversarial Network (GAN) was studied. We collected data from VIS and IR cameras under severe conditions such as snowfall, fog, and heavy rain. The infrared data images from KAIST were used for data learning and verification. We confirmed that the proposed data augmentation method improved the object detection performance, by applying generated dataset to various object detection networks. Based on the study results, we plan on developing object detection technology using only cameras, by creating IR datasets from numerous VIS camera data to be secured in the future and fusion with VIS cameras.

Recently, interest in Prognostics and Health management (PHM) has been increasing as an advanced technology of maintenance. PHM technology is a technology that allows equipment to check its condition and predict failures in advance. To realize PHM technology, it is important to implement artificial intelligence technology that diagnoses failures based on data. Vibration data is often used to diagnose the state of the rotating machine. Additionally, there have been many efforts to convert vibration data into 2D images to apply a convolutional neural network (CNN), which is emerging as a powerful algorithm in the image processing field, to vibration data. In this study, a series of PHM processes for acquiring data from a rotary machine and using it to check the condition of the machine were applied to the rotary table. Additionally, a study was conducted to introduce and compare two methodologies for converting vibration data into 2D images. Finally, a GUI program to implement the PHM process was developed.

Due to recent development of sensor technology and IoT, research is being actively conducted on PHM (Prognostics and Health Management), a methodology that collects equipment or system status information and determines maintenance using diagnosis and prediction techniques. Among various research studies, research on anomaly detection technology that detects abnormalities in assets through data is becoming more important due to the nature of industrial sites where it is difficult to obtain failure data. Conventional machine learning-based and statistical-based models such as PCA, KNN, MD, and iForest involve human intervention in the data preprocessing process. Thus, they are not suitable for time series data. Recently, deep learning-based anomaly detection models with better performances than conventional machine learning models are being developed. In particular, several models with improved performance by fusing time series data with LSTM, AE (Autoencoder), VAE (Variational Auto Encoder), and GAN (Generative Adversarial Network) are attracting attention as anomaly detection models for time series data. In the present study, we present a method that uses Likelihood to improve the evaluation method of existing models.

Tool condition monitoring is one of the key issues in mechanical machining for efficient manufacturing of the parts in several industries. In this study, a tool condition monitoring system for milling was developed using a tri-axial accelerometer, a data acquisition, and signal processing module, and an alexnet as deep learning. Milling experiments were conducted on an aluminum 6061 workpiece. A three-axis accelerometer was installed on a spindle to collect vibration signals in three directions during milling. The image using time-domain, CWT, STFT represented the change in tool wear of X, Y axis directions. Alexnet was modified to learn images of the two directional vibration signals, to predict the tool condition. From an analysis of the results of learning based on the experimental data, the performance of the monitoring system could be significantly improved by the suitable selection of the data image method.