Convolutional Neural Networks Help Detect Data Anomalies
post-template-default,single,single-post,postid-21556,single-format-standard,bridge-core-1.0.4,mega-menu-top-navigation,ajax_fade,page_not_loaded,,qode_grid_1400,qode-content-sidebar-responsive,wpb-js-composer js-comp-ver-5.7,vc_responsive

Anomaly Detection Using Convolutional Neural Networks

Anomaly Detection

Anomaly Detection Using Convolutional Neural Networks

Technically, an anomaly can be considered as an outlier that does not conform to the common pattern or the nature of the class to which it belongs. The definition of anomaly is domain dependent. For example, a zebra in a herd of horses can be considered an anomaly and can therefore be distinguished from the rest of the data.

Many real-world systems exhibit periodic behavior that produce data of a periodic time-series nature. Anomalies can occur within a wide range of systems. Examples would include medical diagnostics, manufacturing assembly lines, robotic ecosystems, oil and gas production, etc. Anomaly detection is the process of detecting behaviors that are outside the predetermined limits established for a particular machine or device. Real-time response to those anomalies is critical to avoiding costly asset downtimes and/or potentially hazardous asset malfunctions. An application designed to automatically detect changes in these periodic patterns or cycles dramatically improves the performance and reliability of critical assets.

Time Series Models for Anomaly Detection

Over the past several years various types of recurrent neural networks (RNNs)– specifically long short-term memory (LSTM)– networks have been employed for real-time analysis and anomaly detection in time series data. Historically, the training of LSTM has been difficult, resulting in limited performance.  Contrastingly, convolutional neural networks (CNNs) have proven successful in the processing of images for classification tasks. The pattern classification strategy utilized for CNN image processing can also be applied to other types of signal processing.

How are multiple periodic signals handled within the CNN framework?

CNN Handling of Multiple Periodic Signals   

Within a real-world system multiple correlated signals (PMCS) are manifested as a multi-dimensional signal where each signal works as a single feature of that PMCS signal. Let’s look at how CNN can be used for such a system.

Data processing: Multiple signals can be evaluated as multiple features within the process of anomaly detection. For successful pattern detection within periodic signals data processing plays a vital role and includes the following steps:

  • Period detection
  • Data normalization—if necessary
  • Data wrangling of parameters (i.e., signal value or duration)
  • Identification and response to missing data
  • Feature selection based on signal correlation (i.e., signal cycle frequency and/or signal changes)
  • Real-time model adjustment for fully automatic execution


CNN Configuration:

  • Multi-channel CNN for multiple signals: Correlated multiple signals can be considered as a multi-dimensional time series signal. A multi-channel deep CNN would be befitting for these correlated signals (PMCS), where each channel in the input layer corresponds to a single feature (dimension) of the time
  • Identifying phases/order of occurrences: Training a data-adapted classifier to detect patterns or sub patterns within the time series data would be an appropriate approach in this scenario. Those patterns and sub patterns underlying the time series over different periods can be segregated into different classes.


Detecting abnormal behavior or anomalies of a time series nature can produce positive impacts for any specific use case. Improving production efficiency and detecting bottlenecks in the production line are just a couple of the potential benefits that can be achieved via this type of analysis. There are positive implications for a wide range of industries and verticals.