Deep learning is a perspective and rapidly developing part of machine learning methods. Deep learning aimed at progressive extraction of higher-level features from the raw input data, therefore effectively “teaching” computers to “learn by example”.

Deep learning applications currently spans from entertainment, visual recognition, and language translations to self-driving cars and healthcare. Here we will briefly discuss how deep learning methods could benefit telematics.

Deep Learning in brief

In a rather broad sense, deep learning could be defined as a set of deep neural network technologies and algorithms, applied to solve various problems. We have already briefly covered some principles behind neural networks in our blog. As being mentioned above, deep learning is actually a branch of machine learning.

Both deep learning and machine learning could deal with regression (numeric) and classification (non-numeric) problems, and both include the following steps:

  • training and data/model testing
  • optimization routines and finding weights
  • further analysis to make the model best fit the data

Normally, in neural networks, input units receive different forms of information that the neural networks will try to learn about.

Output units, located on the opposite side of the network, signal how it responds to the learned information. Hidden units are located in between input and output ones and form the major part of “an artificial brain”.

The connections between units are represented by a weight, which is either a positive or negative number. The higher weight corresponds to a higher influence that one unit has on the other.

The important difference between deep learning and machine learning is the number of hidden layers: in deep learning case, the model being trained contains more than one hidden layer between the input and the output layers. Such multiple processing layers structure allows utilizing complex computational models efficiently and to learn data representations with multiple levels of abstraction.

Based on the short description above, deep learning could be perceived as the evolution of machine learning. It employs a programmable neural network that provides machines the opportunity to make quite accurate decisions without humans assistance.

Deep Learning in Telematics: approaches

In the case of telematics, deep learning at the moment mostly targeting driver behavior analysis, vehicle security,  driver identification, and some related insurance telematics challenges. Further, we will briefly point out a few deep learning approaches to telematics, based on research papers.

In [Abenezer Girma et al., 2019] authors proposed a deep learning model that can identify drivers from their driving behaviors based on vehicle telematics data. Taking into account that OBD-II data is a sequence of sensor data collected over time from the car during a driving trip, authors formulate the driving identification problem as a time-series prediction task where a window sequence of OBD-II data ′S′ from time T start to time T finish can be identified as it is from one of the drivers out of a given set of individuals.

Also, the performance of the model to give a correct prediction under different levels of environmental noise influence and sensor, anomalies have been investigated. Deep LSTM (long short-term memory) Driver Identification model architecture is shown in the figure below.

It was shown that LSTM has an inherent ability to remember temporal information in data and keep it saved for many time steps than the other conventional machine learning approaches. Also, even under increasing noise and outliers effect, the proposed approach maintains its accuracy above the acceptable value, 88%.

In [Jingqiu Guo et al., 2018], authors were aimed at developing an effective approach that can extract the low dimensional high-level features of driving behavior, and accurately explore the hidden behavior sub-groups across a heterogeneous population. They proposed a learning framework to study driving characteristics on GPS data, in an unsupervised feature learning and classification architecture, which is called AESOM Authoencoder Self-Organizing Mapping).

For simplicity’s sake, authors consider GPS data as a unique raw input source. Most of the traffic parameters were found to have mixed effects on the road network. As a result, authors conclude that by extracting highly correlated time-series data of latent features and clustering into driving risk groups, the proposed approach can be an effective tool for proactive road management strategies.

Another approach that can effectively extract high level and interpretable features describing complex driving patterns proposed by [Weishan Dong et al, 2016]. It consists of two components: data transformation and feature learning by deep networks. For simplicity, authors consider GPS data as the only input.

First, the segments of length Ls being generated and the basic features from the raw GPS sequence being calculated. Then, for each segment, authors assign Lf neighboring points into a frame and generate the statistical features.

A major requirement on the data quality for characterizing fine-grained driving behaviors was that the GPS data sampling rate must not be too low. Taking the driver identification as a sample task, experiments on a large real dataset showed that the proposed deep learning approach significantly outperforms traditional machine learning methods as well as the state-of-the-art feature engineering methods that mostly rely on handcrafted driving behavior features.

In mobile telematics case, signals from smartphone sensors could also successfully be utilized to indicate Intricate and diverse patterns, analyze over-time dynamics, and apply deep learning methods to make weighted decisions and classify the whole sequences to certain events, like speeding, turn, phone handling, etc [Wayne Zhang, 2019].

Based on recent research papers and continuously growing interest in machine learning methods, it is pretty clear that telematics also will gain more from deep learning with time. Even currently available methods seem to be very promising in driver identification, security, and driver behavior. With the further advancing of accuracy and robustness of such methods, the possible applications of deep learning in telematics will spread more.

References

  • Weishan Dong et al., Characterizing Driving Styles with Deep Learning, 2016.
  • Wayne Zhang, Applications of Deep Learning in Telematics, 2019.
  • Aston Zhang et al., Dive into Deep Learning, 2020.
  • Michael V. Copeland, Deep Learning Explained, 2016.
  • Abenezer Girma et al., Driver Identification Based on Vehicle Telematics Data using LSTM-Recurrent Neural Network, 2019.
  • Guangyuan Gao, Mario V. Wüthrich, Convolutional Neural Network Classification
    of Telematics Car Driving Data, 2018.
  • Jingqiu Guo et al., Driving Behaviour Style Study with a Hybrid Deep
    Learning Framework Based on GPS Data, 2018.
  • Michael V. Copeland, What’s the Difference Between Artificial Intelligence, Machine Learning and Deep Learning, 2016.
  • http://ictacademy.com.ng/deep-learning-vs-machine-learning-understanding-the-differences
  • https://www.kdnuggets.com/2016/10/deep-learning-key-terms-explained.html
  • https://www.datadriveninvestor.com/deep-learning-explained
  • https://www.zendesk.com/blog/machine-learning-and-deep-learning
  • https://talks.navixy.com