PB4: Contributed Session

Room: Old Main Academic Center 3090

Webex Link

Chair: Abdur Rahman

Abdur Rahman, Mississippi State University

Time: 2:40 pm - 3:00 pm (CST)

Title: Multi-variate Time Series Anomaly Detection through Reinforcement Learning


With the emergence of the smart sensors and wearable technologies, enormous amount of real-time high-dimensional streaming data have been collected. Identifying significant events such as anomalies or outliers has become a key and challenging task using the massive time series data. This research aims to address two challenges: 1) a scarcity of labeled data, and 2) the limitations of the current approaches in modeling multivariate time series. We propose a novel reinforcement learning (RL) framework to detect anomalies in partially labeled multivariate time series data. A proximal policy optimization (PPO) method is applied to optimize the underlying policy for detecting anomalies in partially labeled data. To investigate the impact of different number of available labels to the model, we studied from 5% to 0.1% of training data as labeled. We first constructed a synthetic dataset to theoretically validate the performance of the proposed RL model. Then, three state-of-the-art anomaly detection methods are compared to our proposed RL model on two widely used practical datasets. With its superior performance, this proposed method holds potential as a practical anomaly detection tool for numerous industries.

(Joint work with Junfeng Ma and Haifeng Wang)

Yibin Wang, Mississippi State University

Time:  3:00 pm - 3:20 pm (CST)

Title: Prediction of Aircraft Mission Capability using Expirable Attention-Based Network


Military aircraft datasets are analyzed from a readiness perspective. Readiness can be described as the percentage of fighting force available to perform a mission. The readiness state of Non-Mission Capable (NMC) is critical to identify to achieve mission success before a failure occurs and also in maintenance process. In current readiness state analysis, domain experts must manually assess significant amounts of data and identify both the frequency and severity of failure modes, which is time-consuming and potentially subjective. Additionally, manually input information is difficult to apply conventional learning methods to solve. In this research, we proposed an expirable attention-based deep learning method to predict aircraft mission capability. Expire-span approach is utilized to retain the important information and expire the irrelevant information. This forgetting of memories enables attention mechanism to handle previous time steps efficiently and help the model identify critical information. An EA-6B aircraft dataset with both categorical and manually input data is investigated. The efficiency of the proposed method is analyzed to compare with existing approaches and it is demonstrated that expirable network is applicable with faster training using less memory.

(Joint work with Raed Jaradat and Haifeng Wang)

Timothy Wunrow, Mississippi State University 

Time:  3:40 pm - 4:00 pm (CST)

Title: The Effect of Data Availability and Domain Adaptation for Part Certification in Metal-Based Additive Manufacturing


Part certification is an important task in additive manufacturing (AM) quality assurance. Machine learning has already been extensively applied to AM certification for in-situ anomaly detection. However, AM experiments and data collection are costly, making it expensive to obtain enough data to train a reliable model that can be effectively generalized. To make a larger dataset, data from different systems could be combined. However, each set of data may have a different distribution due to the difference in machine setup and sensing configurations. Domain adaptation (DA) is a technique that can be used to adapt the data from one or more “source domains” to enhance the learning of a different but related “target domain”. Therefore, DA can be used to combine data collected from different AM systems to build more reliable machine learning models for AM anomaly detection. In this research, experimental data are collected from two different metal-based AM part designs for cross-system certification of the direct energy deposition process. The effect of training set size on the anomaly detection prediction accuracy is firstly examined. Furthermore, the two datasets are combined with and without DA to investigate the effects of combining data from different domains on anomaly detection accuracy.

(Joint work with Wenmeng Tian)

Jinhee Yu,  Mississippi State University

Time:  4:00 pm - 4:20 pm (CST)

Title: Speech emotion recognition using convolutional neural networks with self-attention


This paper tests CNN-self-attention models for speech emotion recognition (SER) and compares its performance with several widely studied Convolutional Neural Network (CNN) models from other studies. This work aims to identify which CNN architectures are effective for SER tasks and examine whether the self-attention layer can benefit SER tasks when used with these CNN models. The first step in our approach involves testing some of the popular CNN architectures such as VGG-16, ResNet18, ResNet101, MobileNetV2, DenseNet121, and EfficientNet-B7 on the RAVDESS dataset, a publicly available emotional speech dataset. We then select three of them that achieved the highest accuracy as candidates for the CNN layer of the CNN-self-attention models to evaluate them. We conducted all the experiments using stratified 10-fold cross-validation on RAVDESS, using the MFCC features extraction method. DenseNet121 achieved the highest average accuracy with 80.21% among those CNN architectures, followed by MobileNetV2 at 67.36% and EfficientNet-B7 at 61.46%. When we tested these three models with multi-head self-attention, the results show that all the DensNet121, MobileNetV2, and EfficientNet-B7 models with self-attention have lower accuracy, achieving 62.99%, 58.61%, and 53.19% respectively, than when they are used independently. As a result of our findings, we conclude that merely adding self-attention to CNN architecture in SER does not appear to improve performance.

Lamiae Taoudi, Mississippi State University

Time:  4:20 pm - 4:40 pm (CST)

Title: Crop Yield Prediction Using Machine Learning Techniques


In recent years, precision agriculture has become crucial for the food security and availability of nations. This type of farm management employs various technologies to guarantee profitability, and sustainability. In this study, crop yield data collected from different farm locations in the state of Mississippi is analyzed. The aim is to provide predictive decisions that can help farmers increase the yield of their crops at the lowest cost. Different machine learning algorithms for regression with various pre-processing scenarios are used to find the best algorithms, parameters and solutions. Linear regression with stepwise selection, LASSO penalty, Ridge regression, Random Forest and other methods are used to choose the best combination of variables to model the relationship between different fertilizer nutrients, soil conditions and the crop yield. The results show the ability of the models to predict the crop yield in the region studied.