dc.description.abstract | With the development of smart manufacturing, many machines will be equipped with sensors to anomaly detection. However, anomaly detection has many problems, such as a large number of false positives, difficulty in parameter adjustment, required labeled normal and abnormal data, etc. One of the most common problems is data imbalance. This problem has a big impact on anomaly detection. Not only affecting the training of the model, but also inaccurate analysis results.
Time-series data means data will continue updating, such as sensor. In addition, if there is a large amount of data, and it is difficult to learn the pattern by manually extracting the feature space. Through the deep learning model to transform the data into a new feature space to learn the feature space. There are many studies that have found autoencoder that use unsupervised learning techniques. Because it has the ability to process spatial data through convolutional nerves to detect abnormal behaviors. Deep learning apply in image, text and classification. However, in this research is still an unexplored field.
This study proposes the concept of pre-training using autoencoder to dealing with data imbalance. Because autoencoder is an unsupervised learning neural network. Therefore, the autoencoder will try to find the best weight to training model. Therefore, the time series is reconstructed through the encoder and decoder. With the ability of long and LSTM to learn long-term data. It is suitable for time-series forecasting or anomaly detection. The trained Autoencoder-LSTM model can reconstruct time series data, and the neural network can effectively predict periodic time series data. Therefore, the goal of our research is using an autoencoder-long short-term memory network to solve the data imbalance. Let the model performance can effectively detect abnormal data and provide a new perspective to solve the data imbalance classification problem. | en_US |