Search

In recent years, convolutional neural networks (CNN) have become leading algorithms for many computer vision tasks. Many studies have used deep learning and artificial intelligence (AI). This paper summarizes the development of deep learning approaches and the impact of AI on marketing and customer behavior and discusses the challenges facing marketing research. The ma jor marketing application of AI is to support customer decision-making. However, most previous marketing research has focused on how to contribute to marketing decision-making in practice. We believe that there is room for research on how deep learning approaches can be used to support marketing decision-making. One area that has seen major developments in deep learning approaches is that of image data analysis, including video analysis. In marketing decision-making, market segmentation and positioning strategies are closely related to image data analysis. We believe that more research based on deep learning approaches should be conducted to support decision-making in these areas. Therefore, to support decision-making in marketing, we present studies on market segmentation and positioning strategy formulation as examples of studies that combine deep learning approaches with traditional marketing research. [ABSTRACT FROM AUTHOR]

We have shown that the shape of parking vehicles and scenes can be estimated with high accuracy using a high-resolution millimeter-wave radar for automated parking and parking assistance in our previous studies. In this paper, we propose a method for estimating the shape of parking vehicles from only range and Doppler information, using a low-cost millimeter-wave radar without multiple antennas by applying deep learning techniques. It is shown that it can be estimated with high accuracy in real-world parking scenes. [ABSTRACT FROM AUTHOR]

Visual illusions are phenomena in which the actual physical parameters of an object are out of alignment with visual perception, such as when an object that is not moving appears to move, or when the color, brightness, or size of an object appears to be different from its reality. It is also the perception of such discrepancies. Illusions have attracted many people for a long time because they are surprising to the viewer and have a strong entertainment value. Illusions are thought to express the characteristics of visual information processing in the eye and brain, and are an important tool in the study of vision, and the field of research has been developing remarkably. In recent years, research methods have actively incorporated not only psychophysical methods but also modern brain science methodologies such as fMRI, and more recently, research utilizing artificial intelligence has also progressed. This paper focuses on the study of visual illusions using deep neural networks (DNN) which is being conducted in the authors laboratory and is becoming popular worldwide. [ABSTRACT FROM AUTHOR]

To understand the experience of seeing is difficult and is being pursued day and night, especially in psychology and neuroscience. However, there are hard areas to research, such as estimating the structure of receptive fields (RF) in the higher-order visual cortex with humans and animals as research targets. Deep neural networks (DNNs) are being reported that have similar properties to visual neurons and the possibility of using DNNs as alternative re- search targets to the biological brain has emerged. Therefore, in this paper, I discuss whether DNNs can be our re- search subject. In this research, I applied the reverse correlation method, which has revealed RF of visual neurons, to DNNs to estimate RF of units in well-trained VGG-16. As a result, the properties of the RF of VGG-16 units were similar to visual cortex neurons. The result suggested that DNNs may be a good alternative model for our research, but also suggested limitations of the method. To solve remaining problems, psychology research that develop the better methods and deep learning research that provides better alternative models must go hand in hand. [ABSTRACT FROM AUTHOR]

This paper reports a method to estimate inflow wind conditions undisturbed by wind turbine rotors while using a deep learning technique. For creating the inflow wind estimation system, datasets taken from nacelle-mounted anemometers and a vertical profiling Light Detection and Ranging (LiDAR) were employed. Our previous study showed that unreasonable errors in wind direction observations were enhanced by applying the deep learning model. However, a large amount of data was required to train the deep learning model from scratch. Therefore, long-term measurement with doppler LiDAR is required, and its cost tends to be high. In this study, we propose a transfer learning method to ensure the performance for other wind turbine with only short-term data. The proposed method was evaluated on the actual data, and the results showed that it was possible to estimate the trend of doppler LiDAR wind direction. Future work includes the evaluation of different condition sites such as the adaptation from flatland to mountainous sites. [ABSTRACT FROM AUTHOR]

Operant conditioning has not been well adapted to social behavior despite a growing interest among behavior analysts. This is due in part to technological limitations in measuring social behavior, especially in real time. This paper presents a summary of three of the present author􀊟s attempts to adapt computer vision technologies to the study of social operant behavior of zebrafish (Danio rerio). Experiment 1 examined whether the location of a single zebrafish in a three-dimensional (3D) space could be tracked in real time. In Experiment 2, locations of two zebrafish were tracked simultaneously. A transparent partition was placed between the two fish to aid in their identification. A food reinforcer was delivered when the two fish approached one another across the partition. Experiment 3 examined whether it is possible to track multiple fish without such a partition. Results of these three experiments were promising, suggesting that computer vision can be useful in the study of social operant behavior. [ABSTRACT FROM AUTHOR]

This paper focuses on the automation of work measurement using a camera to capture workers’ hands in an assembly factory and analyzing the image to estimate the work content. In recent years, the technology to collect, analyze, and visualize manufacturing data from various devices installed in processing machines has been improving. However, it has not yet reached the point where workers’ manual operations can be analyzed and visualized in real time. In this study, it is proposed that high estimate accuracy can be achieved using two analysis procedures to analyze a worker’s image with deep learning: setting the region of interest and estimating the work content from the features in the region of interest. As a result of an evaluation experiment conducted at an actual factory using the proposed method, 99.5% of the work was correctly estimated. Based on these results, the method proposed in this study is now being examined in anticipation of introducing it for practical application. [ABSTRACT FROM AUTHOR]

A Faster and data-driven method by deep learning, compare with a traditional method by FEM, is proposed for thermal simulation of Wire Arc Additive Manufacturing （WAAM）. In this paper, cross sectional temperature distributions of WAAM-made parts are predicted and shows a good agreement with FEM method. [ABSTRACT FROM AUTHOR]

In the conventional yaw control, the wind direction measurements obtained from sensors installed on the wind turbine nacelle are used. However, the wind direction based on this primitive approach would result in a yaw misalignment, because the wind direction taken by the nacelle-mounted sensors are strongly influenced by the rotor rotation. To mitigate the yaw misalignment, we have developed a new system to estimate inflow into a wind turbine based on an AI (Artificial Intelligence) technique with DNN (Deep Neural Network). In this research, we implemented this new AI-based technique on an actual wind turbine control system by adding an external module. This paper reports the impact of using the AI-based inflow estimation system for wind turbine yaw control from an experiment using an upwind wind turbine. As a result, the annual energy production of the wind turbine using the system was assumed to increase. [ABSTRACT FROM AUTHOR]

Building energy simulation (BES) is commonly conducted in architecture design process to evaluate building energy performance. The coupling between computational fluid dynamics (CFD) and BES not only improves energy simulation accuracy but also makes it possible to simultaneously consider energy consumption and indoor environment. Nevertheless, to conduct high fidelity CFD simulation is generally time-consumed and thus it is almost impractical to carry out a long-term coupled simulation that requires multiple CFD executions. A fast and accurate prediction method is therefore required to serve as a surrogate for CFD in the coupled simulation. Inspired by successful applications of deep learning neural networks (NNs) in various fields due to the high computation speed and prediction accuracy, the authors proposed a deep learning-based prediction method to achieve fast and accurate prediction of indoor air distributions. This paper provides a general introduction to the proposed prediction method with regard to its principle, implementation, performance, and application. Predictions of two-dimensional isothermal flow and three-dimensional non-isothermal flow are demonstrated. The results confirmed the feasibility of NN models for fast and accurate indoor airflow prediction. Meanwhile, as an example of practical application, the NN model is coupled with a BES tool to implement a coupled simulation framework for fast energy simulation considering non-uniform indoor environment. The coupling scheme is introduced and validation of the crucial functionality of the framework is presented. [ABSTRACT FROM AUTHOR]

Overhead crane is a product used to transport heavy objects in a factory, and advanced operating skill is required to operate safely and with less equipment failure. However, the shortage of skilled operators has been pointed out in recent years. The purpose of this paper is to develop a method for estimating tense state of rope and weight of suspended load at the same time with the aim of improving the safety of overhead cranes without relying on operator skills. When estimating multiple states using independent estimation models, the issue is the lack of computer storage space and calculation speed due to the increase of the parameters. Therefore, the authors adopted a method that simultaneously estimates several states in a single model. To estimate the tense state of rope and weight of suspended load from the driving signals of the induction motor mounted on the overhead crane, a multi-task neural network that can express the nonlinearity of motor characteristic was used. The teaching data used for learning network parameters was created by using object detection technology to quantify the transportation of markers attached on load block and suspended load. As a result, it was clarified that the rope tension state and the load can be estimated simultaneously by inputting the driving signals of the motor to the multi-task neural network. Moreover, it was shown that the rope tension probability and the estimated load increased synchronously as the motor signals change, and the tension detection was output.

The mechanical properties, such as Vickers hardness and crack length, of various WC-FeAl hard materials were predicted using deep learning via a convolutional neural network (CNN) trained on microstructures. The accuracy of the predictions was verified using gradient-weighted class activation mapping (Grad-CAM), which is a kind of image visualization technology that identifies important structural features for AI classification based on mechanical properties. The accuracies, expressed as coefficients of determination for unknown samples (test data), were found to be 0.89 and 0.75 at most for Vickers hardness and crack length, respectively. The AI correctly recognized microstructural quality and determined classes that represented differences in mechanical properties as evidenced by the feature maps obtained, indicating that CNN prediction was a powerful tool for analyzing WC-FeAl hard materials. [ABSTRACT FROM AUTHOR]

Nowadays, along with the popularity of E-Commerce, the marketing strategy of retail stores has been more complicated with O2O or Omni-channel. Therefore, Customer Relationship Management (CRM) is one of the important issue for the retail stores. It can be difficult to predict customers future behavior with the simple quantitive information such as purchase frequency since each customers are widely diversified. Although the company can obtain the variety of customers information from their online activity, the use of access history is still limited. In this paper, we defined “the variety of user access patterns” collected from their web browsing history and it shows the patterns they visit the website. Finally, we verified its effectiveness with developing a DNN model to predict customers future behavior. [ABSTRACT FROM AUTHOR]

Research and development are actively conducted for recent advances in the self-driving and unmanned vehicles, along with further efficiency and safety. In such fields, obstacle avoidance is frequently addressed in control applications through the concept of active safety technology. This paper investigates the real-time optimal obstacle avoidance control that the optimal path generator embedded model predictive controller. By offline computation, a large number of optimal obstacle avoidance control solutions are calculated to generate the supervised learning data for the deep-learning neural network used for the optimal path generator. Then, constructed optimal path generator is embedded in the model predictive controller to generate the reference command for the tracking task during the optimization process. The utility of the proposed controller is evaluated by both the numerical simulation and experimental testing of the obstacle avoidance maneuver of the micro-unmanned vehicle. Results show that the proposed real-time optimal controller successfully performs obstacle avoidance.

In this paper, the prediction of crack propagation with two cracks using machine learning is described. The analysis results of crack propagation by s-version FEM (s-FEM), which combines the automatic mesh generation technique, are used for generation of training and validation datasets. Plural crack propagation with the different vertical distance between the two cracks as a variable are analyzed. The analysis cases are divided into training and validation datasets. In training process, the input parameters are the coordinates of 4 crack tip, the output data are crack propagation vectors, the number of cycles for crack propagation of 0.25 mm. Initial crack configurations should be specified. After the specification, the predictor iteratively predicts crack propagation direction and the number of loading cycles. A prediction accuracy depends on the training datasets, which contains 0.25 mm length of each crack propagation in this study. To improve prediction accuracy, the data augmentation is effectively applied. In case of plural crack interaction, when the crack tips close each other, the accuracy gets worse and worse. Reducing datasets which satisfy the crack coalescence condition, it is shown that the prediction accuracy is improved. Even if training datasets are not enough number for accurate prediction, it is shown that the prediction accuracy is improved by the data augmentation.

Various fields, such as the paper industry, chemical engineering, and renewable energy, are faced with gas-liquid two-phase flows and are being studied by visualization and observation. Although it is necessary to quantitatively evaluate the characteristics of bubbles, there is a limitation in the amount of labor required for detection and measurement by human observation of images. There are no examples for bubbles in polymer electrolyte membrane water electrolysis (PEMWE), where the bubbles in PEMWE have heterogeneous backgrounds, unpatterned patterns, and unclear bubble contours. Existing methods for detecting these bubbles are not expected to be accurate enough. In this study, a deep learning-based bubble detection method using convolutional neural networks (CNN) was developed for bubbles in PEMWE. Our method has two novel approaches: first, we developed an algorithm that automatically draws a pseudo-bubble image based on an actual bubble image on an arbitrary background as a method for automatically generating training data for the CNNs. Second, we developed a CNN-based bubble detection method that uses the motion of bubbles, specifically, the difference between the bubble image and the image from one frame ago, as input. Finally, we tested the algorithm on bubble images in a PEMWE and achieved an F1 score of 0.83 or better for all current densities of 0.5, 1, 2, and 3 A/cm2, and an F1 score of 0.844 for the entire detection.

This paper presents a method for predicting the crack growth shape and the number of cycles of a two-dimensional fatigue crack under cyclic loading using a convolutional neural network. All of data sets for train are generated by s-version FEM for fatigue crack propagation analysis. The crack propagation simulations were simulated with different slant angles. Crack tip coordinates, crack growth vectors, and numbers of cycles are prepared as a set of train data for one prediction step, which is determined by the minimum mesh size of the crack tip in the s-FEM simulation. Data augmentation technique, which adds a slight noise to input data, is introduced as regularization in this work. We'd like to evaluate the effectiveness of the data augmentation. Additionally, the interpolation ability and extrapolation ability of the prediction model are evaluated. The crack growth shapes and the number of cycles in the prediction step can be predicted within 6%, 11.1% difference with the reference.

Since Deep Convolutional Neural networks (DCNs) have achieved excellent general object recognition performance, researchers in the fields of systems neuroscience and visual psychophysics have attempted to understand visual processing in the brain via DCNs. In this talk, I will review the progress of DCNs in computer vision, and introduce the authors' visual neuroscience studies using DCNs. Then, I will discuss the prospects for multi-omics analysis of psychonomic science data based on recent research trends in deep learning and multimodal processing. [ABSTRACT FROM AUTHOR]

Monitoring structural integrity has been demanded to achieve a sustainable society against disasters, including seismic and extreme wind events, and thus Structural Health Monitoring (SHM) system is one of the significant technologies. The natural frequency and the yield displacement of structures will be the significant damage indices to assess the structural integrity. However, they would contain variances to the design values. The structural integrity assessments using the SHM should allow for the uncertainties in structural specifications, such as the yield displacement, to improve the assessment accuracy. Additionally, the earthquake ground motion will have uncertainties regarding the acceleration levels, spectral characteristics, and others. These uncertainties can lead to assessment difficulty, reducing the accuracy of the assessment. The authors propose the method using the deep Convolutional Neural Network to overcome this problem. This method using CNN has two features. The first is to suggest using the response spectra as the training data of CNN. The second is to create a simple CNN architecture with a high classifier accuracy using Bayesian optimization. This paper demonstrates that the image features detected by CNN enable us to accurately assess the structural integrity even when uncertainties in structural specifications and seismic motions are superimposed.

This paper investigates the utility of nonlinear optimal feedback control for a rotary pendulum based on deep learning. The neural network trained with supervised learning using solutions of the minimum-time optimal control problem provides the feedback controller, which generates control variables by current state variables. The utility of the proposed idea is verified by the numerical simulation and experiment considering the pendulum angular velocity constraint. Results demonstrate the feasibility of commanding the rotary pendulum with the proposed nonlinear feedback controller.

This paper presents a real-time self-attitude estimation method which utilizes the clues to the direction of the gravity hidden in images and structures of the environments. In the proposed method, the angular velocity is integrated using a gyroscope, a camera-based method estimates the gravity direction, and a LiDAR-based one also estimates the gravity direction, respectively. These estimations are integrated using the EKF (extended Kalman filter). The camera-based gravity direction estimation uses a DNN (deep neural network) which learns the regularity between the gravity direction and the landscape information. By learning the regularity, the proposed DNN can infer the gravity direction from only a single shot image. The DNN outputs the mean and variance to express uncertainty of the inference. The LiDAR-based gravity direction estimation extracts vertical planes from the surrounding environment measured by the LiDAR, and outputs the gravity direction based on their normals. By using both the camera and the LiDAR, more robust and accurate estimation can be achieved. To show the DNN can estimate the direction of gravity with uncertainty expression, static validations on test datasets are performed. Dynamic validations are also performed to show the proposed EKF-based method can estimate the attitude in real time. These validations are performed in both simulator and real world to compare the proposed method with conventional methods.

In this paper, we propose a method to predict the failure of factory equipment by machine learning architectures using vibration data. We design the model so that we can predict robustly the failure of the equipment in advance. We use a Gaussian Mixture Model (GMM), a machine learning architecture, to calculate abnormality value which is used for the decision whether the state of the equipment is normal or abnormal by thresholding. We also use Long Short-Term Memory Autoencoder (LSTM-AE), one of the structures of the deep learning algorithm, for feature extraction. LSTM-AE model learns both spatial and temporal patterns which are difficult to capture with conventional machine learning algorithms. We conducted the prediction experiment using vibration data obtained from actual mechanical equipment, to confirm our method can predict the failure more robust than conventional methods. From this experiment, we found that the abnormality value tended to exceed a threshold value before the actual failure, indicating that the failure can be predicted in advance by our method. Besides, when compared with conventional methods, we found that the transition of abnormality and the accuracy of failure prediction were almost the same in all cases, but we also showed that the proposed method has superiority on robustness compared to conventional methods about the transition of abnormality and the setting of the threshold.

Computer Aided Engineering (CAE) is indispensable for vehicle design to reduce the development cost; however, its computation time is a heavy burden when tuning design parameters. In this respect, several studies have been carried out for replacing CAE with machine learning-based surrogate models. In this paper, we propose a novel neural sequence network-based surrogate model for CAE using Recurrent Neural Networks (RNNs), which are neural networks that treat sequences such as temporal sequences. Our target task is the NCAP Fishhook test to evaluate vehicle dynamics of the rollover propensity. We propose a machine learning model with a sequential model to calculate the response of the NCAP Fishhook test from vehicle parameters such as tire and suspension characteristics. Our model reduced the error in approximately 10% for the NCAP Fishhook test dataset, which is generated with CAE, compared to that of the baseline neural network model with multi-layer perceptrons (MLPs). Furthermore, to improve performance and stability, our model has the following task-specific characteristics: (1) the skip connection, (2) the hybrid loss, and (3) the scheduled sampling. We confirmed that the skip connection reduced errors in the additional ablation study. Our experiments showed that the sequential model is effective as a surrogate model for CAE, and we also find that there is still room for the improvement regarding the dataset and the model because the accuracy is not saturated with the current dataset.

Fall is one of the main factors causing the extension of the patient’s hospital stay. In some cases, fall leads the elderly patients to bone fracture, bedridden, and care state. Also in other cases, fall leads to fear, withdraw, and bedridden. However, the medical staff can’t keep eye on all the patients at all the time. Moreover, patient's condition and their environment change from moment to moment. Therefore, the primary screening is need to find the high risk patients and to focus attention on them. Furthermore, real-time risk assessment is need to prevent the patients from falling. In this paper, we propose and develop an automatic real-time fall risk assessment using deep learning as a primary screening of the patients in the hospital. By using our method, we construct a part of the nursing system to reduce the fall incidents and accidents of the patients, and the caregiving load of the medical staff. At the reception, proposed method recognizes patient’s appearance and estimates risks by classifying the fall risk levels according to a hospital’s clinical safety standard, and it stores those safety related information in the electric medical records of patients. While the patients stay in the hospital, medical staffs see the potential risks in the condition of the patients and their environment online through the wearable devices such as smart glasses, mobile terminals, and assistive mobile robots. We prepared a datasets including 16,000 images which collected from public datasets such as MIT places, educational video, actual incident location, and annotated them by ourselves for the clinical safety. The result show a fundamental experiment comparing with manual method by medical staffs and our proposed method surpasses real-time performance and coverage. These automatic fall risk assessment method using deep learning is novel to in clinical safety field.