Download Config Anom
CLICK HERE ===== https://urlin.us/2tlIqI
Download Config Anom
Regarding different compression techniques suitable for time series data, two main categories can be identified: lossless and lossy [8]. On the one hand, lossless compression means that the decompressed data are identical to the compressed data, which tends to produce low compression ratios, e.g., 50% compression. On the other hand, lossy compression techniques are intended to produce a trade-off between the accuracy of the reconstructed data and higher compression ratios, e.g., 90% compression. For the latter category, the use of deep learning (DL) methods has attracted special interest from the research community. Among different DL techniques, autoencoders (AE) represent a promising opportunity in the field of lossy compression. Basically, AEs are a type of deep neural network that has an encoder and a decoder part. The encoder part compresses input data into a number of representations called latent features which have a size much smaller than the input dimension. These representations form the latent space (LS) of the AE. The encoder is usually able to compress the data by discarding non-relevant parts of the data while keeping only the parts that can be effectively used for reconstruction in the decoder part. Another advantage of using AEs is their intrinsic ability for anomaly detection [9]. Note that this allows performing not only compression but also data analysis at the source before transport. Thus, locally distributed data analysis can be performed and used to add more intelligence to the monitoring system, e.g., increasing a nominal monitoring sampling rate when an anomaly is detected [10].
In this paper, we propose a novel method for the lossy compression of time series data using deep AEs along with two methods for anomaly detection that operate on both single and multiple time series. For the compression, instead of compressing the input data using a single AE, a pool of AEs with a different number of latent features is used. Thus, the Adaptive AE-based Compression (AAC) method is presented as an autonomous process that is able to choose the best AE in the pool, i.e., the one that reaches a target reconstruction error with the minimum LS size. The variability of the number of latent features means that the size of the compressed data is not fixed, which draws similarities between AE-based compression and conventional compression methods in which the characteristics of the input data play an important role in the compression ratio. It also means that the compression is adaptive to the variations in the data and hence, compression size is indeed a variable that can be analyzed as additional and extended information of collected monitoring data. It is worth mentioning that AEs are trained using data from the specific sensor/s that they operate. However, since this may not be available from the beginning of sensor operation, generic AEs with moderated compression rates trained for heterogeneous sensor data are used until enough data are collected to train the specific AEs.
Regarding the anomaly detection part, the first method, called Single Sensor Anomaly Detection (SS-AD), takes advantage of the specificity of the trained AEs to detect when the collected data contains an anomaly, e.g., if the sensor malfunctions or some kind of additional noise is introduced to the data from an external source. The second method, called the Multiple Sensor Anomalous Group Diagnosis (MS-AGD), detects anomalies that can affect several sensors in a correlated way, even when they cannot be detected by SS-AD in an independent time series. It does this by comparing data points with a certain degree of reconstruction error values across all the time series involved, making it able to detect subtle correlated anomalies.
Figure 1b provides a deeper insight into the hierarchical architecture needed to run the proposed telemetry data compression and analysis. The first level is at the sensor layer where data ar
https://www.ishan13.com/group/ishan13-qun-zu/discussion/9e9d3fd0-5eef-4caa-8616-855cac8da6b0