Multi-head CNN–RNN for multi-time series anomaly detection: An industrial case study

TLDR: This work proposes a deep learning based approach for supervised multi-time series anomaly detection that combines a Convolutional Neural Network and a Recurrent Neural Network in different ways and refers to this architecture as Multi-head CNN–RNN.

About: This article is published in Neurocomputing.The article was published on 2019-07-19 and is currently open access. It has received 196 citations till now. The article focuses on the topics: Anomaly detection & Deep learning.

Figures

Figure 6: Training time of both base models and TKB models. They are all grouped by the number of sensors used.

Table II: Parameter specification for all the architectures employed in this experimentation.

Figure 2: Layer configuration of all the architectures. Each box contains the name of the layer and its corresponding output shape. Note that, for input layers, the last dimension of the output shape refers to the number of channels (sensors), while for convolutional layers it refers to the number of filters.

Figure 1: Multi-head CNN-RNN architecture for multi-time series anomaly detection. From the left, data coming from sensors are individually processed by independent convolutional heads by means of a window W of length WL. The window slides over the time series with a step of size WS . A feature map F n w is obtained as a result of applying a CNN to the window w of sensor n. Feature maps corresponding to the same window w are then concatenated. Once all windows of all sensor data are processed, the recurrent layers yield the classification outcome by processing all the windows chronologically.

Table VIII: Average Friedman rankings and APVs using Holm’s procedure in g mean for Multi-head architectures. The horizontal dashed line delimits the architectures rejected (located below the line) as a consequence of setting the level of significance to α = 0.05

Table VII: Average Friedman rankings and APVs using Holm’s procedure in g mean for all the architectures. The horizontal dashed line delimits the architectures rejected (located below the line) as a consequence of setting the level of significance to α = 0.05

Frequently asked questions

Q1. What contributions have the authors mentioned in the paper "Multi-head cnn-rnn for multi-time series anomaly detection: an industrial case study" ?

The complexity of this task increases when multiple heterogeneous sensors provide information of different nature, scales and frequencies from the same machine. In this work, the authors propose a deep learning based approach for supervised multitime series anomaly detection that combines a Convolutional Neural Network ( CNN ) and a Recurrent Neural Network ( RNN ) in different ways. The proposed architecture is assessed against a real industrial case study, provided by an industrial partner, where a service elevator is monitored. Within this case study, three type of anomalies are considered: point, context-specific, and collective. The experimental results show that the proposed architecture is suitable for multi-time series anomaly detection as it obtained promising results on the real industrial scenario. 

Q2. What future works have the authors mentioned in the paper "Multi-head cnn-rnn for multi-time series anomaly detection: an industrial case study" ?

However, further research must be done to improve the performance and the required training time of the models generated by this method. Little research has been conducted in in this field and in TL techniques for time series and thus it would be a good line of research for the future. Thus, further research must be done to analyze the performance of the proposed architecture at the time of processing time series with different frequencies. However, the methodology to mask padded values in one-dimensional convolutions is an ongoing research and therefore, it is another line of research for the future.