Abstracts Track 2022

Area 1 - Artificial Intelligence

Nr:	20
Title:	AE based Reconstruction Error Correction using Contrastive Learning for Anomaly Detection
Authors:	Yeojin Jeon and Jun-Geol Baek
Abstract:	Anomaly detection is an important task in monitoring and controlling yield reduction and quality deterioration in the manufacturing process. The recent increase in complexity of manufacturing process has led to more diverse types of normal data. Autoencoder(AE), a generative model for anomaly detection, detects the abnormality of input data based on learned representation from normal data. However, we observe that anomaly detection performance of AE drops rapidly on data that contains rare normal data types. In this paper, we propose a novel anomaly detection method. The proposed method corrects AE reconstruction error to learn more various representations of normal data types with self-supervised representation learning. By applying cluster-conditioned detection on feature space, we correct the reconstruction error of AE by increasing the difference in reconstruction error between normal and abnormal data. Experiment results show that the proposed method achieves improvement on anomaly detection performance compared to baseline AE.

Nr:	21
Title:	Multivariate to Multivariate Time Series Forecasting using SACAE and TCN for Manufacturing Machinery Dataset
Authors:	Woo Young Hwang and Jun-Geol Baek
Abstract:	In the manufacturing process, data is collected in the form of correlated sequences. Multivariate to multivariate time series (MMTS) forecasting problem is an important factor in manufacturing. MMTS forecasting is a notoriously challenging task considering the need for incorporating both non-linear correlations between variables (inter-relationships) and temporal relationships of each univariate time series (intra-relationships) while forecasting future time steps of each univariate time series (UTS) data. However, previous works uses deep learning models suited for low-dimensional data. These models are insufficient to model high-dimensional relationships inherent in multivariate time series (MTS) data. Existing MTS forecasting studies are also focused on predicting a single target variable from multiple input variables. It is less productive and time consuming. Thus, we propose two phase multivariate time series forecasting. First, proposed method learns the non-linear correlations between UTS (inter-relationship) through self-attention based convolutional autoencoder. Second, it learns the temporal relationships (intra-relationships) of MTS data through temporal convolutional network and forecasts multiple target outputs. As an end-to-end model, the proposed method proves to be more efficient and derive excellent experimental results.

Nr:	25
Title:	GAF-based Time Series Imaging Transformation Applying Angle of Intersections Matrix for Time Series Classification
Authors:	Yong-Gon Jung and Jun-Geol Baek
Abstract:	A subcategory of time series analysis, time series classification is a task to allocate time series patterns to specific categories. It is a difficult problem because the number of potential temporal features useful for classification is large. Previous studies used 1NN-DTW to measure and cluster the similarity between two time series, or to transform the time series into a two-dimensional image through Fourier or Wavelet transform and apply it to a deep learning model for classification. However, there are some disadvantages in that the temporal information disappears, and the performance fluctuates significantly depending on the basis function or basis wavelet region. In this study, we propose an Angle of Intersection Matrix that improves classification performance by adding the proposed matrix to the existing GAF method in a way that can extract additional information from time series data. Angles in the time interval uses the angle of intersection between the time points to determine how much change has occurred. Directionality using the sign of the slope can have two cases, a positive case and a negative case, only with the angle calculated in the previous step, so the time process is expressed through the slope. The proposed method proceeds in three steps. First, the proposed method calculates the intersection angle of three consecutive time series points from a one-dimensional time series plot in the Cartesian coordinate system. Second, a real matrix is created by the calculated intersection angles and applied to the existing GAF. Third, the proposed method performs a time series classification using the generated AIM-GAF as the input of CNN. To evaluate the proposed method, multivariate time series data with label information was used for more accurate time series classification. These two datasets used in the experiment are open benchmark datasets provided by Olszewski. The wafer dataset is data from six vacuum chamber sensors used to monitor the fabrication of semiconductor microelectronics. The ECG dataset is a time series in which one heartbeat is recorded using electrodes and aims to classify normal (health) and abnormal (heart disease) using the data. The two datasets are commonly labeled with two classes, normal and abnormal, and the length of the data is not the same. To avoid the overfitting problem, 5-fold cross validation was applied and five independent experiments were conducted by configuring different datasets. As each fold completes learning, we reset the weights. The proposed AIM-GAF method showed the lowest error rate compared to previous studies. In this regard, AIM-GAF achieved an error rate of 0.08% for the wafer dataset and 2.50% for the ECG dataset, obtaining the best results. As a result of overall comparison, a decrease in the error rate of at least 0.02%p and a maximum of 3.65%p can be confirmed and on average, the error rate decreased by 1.75%p, resulting in an improvement in performance. The proposed method achieves very competitive results when compared to the existing GAF-based classification method, and it shows improved classification performance compared to existing GAF-based studies in identifying and classifying abnormal.

Nr:	26
Title:	Supply Chain Coalitions within Resource-Constrained Economic Environments: a River Sharing Approach
Authors:	Florina L. Covaci
Abstract:	Companies are continuously confronted by challenges that could impact future resource availability, given the rising complexity of global supply markets. Population expansion and economic development are expected to increase overall resource consumption, perhaps resulting in a scarcity of resources to meet future demand. Reality shows that companies must address not only potential future resource scarcity difficulties (e.g., as a result of a government policy change), but also the fact that scarcity conditions are frequently unpredictable (e.g., will a new policy affect supply, and if so, when?). For example, as the demand for electric vehicles grows, the raw minerals necessary to produce electric batteries, such as cobalt and lithium, are predicted to become scarce. Alternative materials may become available, however these basic materials are also employed in the manufacture of other gadgets. Supply chains in economic environments with limited resources can be evaluated by comparing to rivers with a limited amount of water. Cooperative game theory is frequently used in water resource economics to analyze water resource allocation. In the theoretical literature, this is referred to as the "river sharing problem." The current paper assumes that when a group of players considers breaking away from the rest of society, they are unsure of the partition that the players outside S will form. As a result, on the set of all possible partitions, they assign various probability distributions. These probabilistic beliefs do not always correspond to the behavior of outsiders, beliefs do not have to be consistent with actual choices. Given the beliefs, regardless of how they emerge, one can compute the expected value of S and define the core of the resulting cooperative game. We consider agreements with a single supplier to share scarce resources. Along the supply chain, a number of agents extract the quantity of the scarce resource for use in the production of their own products. Agents each have their own method of evaluating scarce resources, with some having greater needs and higher marginal utility than others. Concave and single-peaked benefit functions are used to represent these heterogeneous valuations, with the peak consumption corresponding to an agent's satiation point. The satiation point is depicted, at which over consumption may result in an increase in storage costs. If the marginal benefits for agents located further downstream are higher, it may be profitable for a coalition to transfer some quantity from one component to another. As a result, the value of a coalition may be greater than the sum of its constituents' values. However, it may be profitable for the agents outside of S to pass some resources from one component to the next, leaving some resources for the agents in S to consume. As a result, the value of a coalition S is determined by both its components and the behavior of the agents outside of S. In other words, the behavior of the agents outside of S has an effect on the value of the coalition S.