Trans RINA, Vol 157, Part A3, Intl J Maritime Eng, Jul-Sep 2015
comes to imprecise data and uncertain information [22]. NNs can perform with non linear data, it is learns and does not need to be reprogrammed. NNs also
Support Vector Machines (SVMs)
SVMs are supervised learning methods using the binary classification that needs some
pre-knowledge before
classification. SVMs map the training data, which is consist of nonlinear data through the kernel function [23]. The role of kernel function is to induce such a feature space by implicity mapping the training data into higher dimensional space where the data is linear separable.
Similar to the NNs, the SVMs methods are capable of learning arbitrary complex regions in the input feature space. A key parameter of the SVMs is the kernel function, such as polynomial or Gaussian, which maps input data to high dimensional feature space where data can be perfectly or close to linear separated. A study done by Li
et.al., [3]
implemented SVMs to analyze
vessel behaviour at a higher level of abstraction.
Bayesian Networks (BNs)
The BNs is a machine learning technique based on the probability that represents a set of random variables and their conditional graph (DAG).
independencies via directed acyclic
Previous studies have been done using the BNs, in which it is applied for anomaly detection on maritime domain and some other domains [7, 24, 25, 31]. Helldin and Riveiro (2009) [31] use the BNs in an anomaly detection research study. The AIS is used as the input data. The study focuses on how reasoning capabilities of the BNs can assist the operator in the control room.
A study done by Mascaro et. al. (2013) [26] defined the advantages of BNs and disadvantages of both the NNs and SVMs approaches. The BNs potentially have two substantial advantages: (1) The models are easy to understand by lay people (including the operator in the control room or other domain experts) and (2) They can include knowledge from the experts as input to the BNs model [26]. On the other hand, the NNs and SVMs do not present a transparent model to the user. Therefore, it is complicated for users like the operator in the control room to understand, interact and explore the model.
2.1 (b) Statistical Methods
Some studies using the statistical or probabilistic approach model have been done, e.g. the hidden Markov model (HMM) [32], Gaussian mixture model (GMM) [14, 33], and adaptive kernel density estimator (KDE) [14, 34].
There are two types of statistical technique, parametric and non-parametric. In parametric techniques, when the data correspond to a particular statistical model, anomaly can be detected rapidly and without supervision. With the non- parametric techniques, no assumption is made about the underlying distribution of data. Although more resources are required to develop them, these methods are effective for automated anomaly detection.
2.2 KNOWLEDGE-DRIVEN TECHNIQUE
In order to partially replace domain experts with the computer system, one must emulate the expert’s capabilities. The knowledge-driven approach should be constructed to provide computer systems that can reason, communicate and interact. The more
complete the
knowledge support, the greater the ability to understand the situation and provide support for the process of anomaly detection.
There are several studies on knowledge-driven
techniques for anomaly detection systems with different techniques such as rule based and description logic [15, 16]. However, in a previous study [7, 24] the hybrid approach was proposed where the expert’s knowledge together with a data-driven approach are combined.
3. OVERVIEW OF OUR APPROACH
According to [23], the BNs is a directed acyclic graph (DAG) comprising a set of nodes and edges which represent the probabilistic dependencies among variables. The nodes with direct edges to other nodes are called the “parent” nodes. However, the nodes with edge pointing into them are known as the “child” nodes. A good example is found in the probabilistic relation between season and temperature. Given the temperature, the network can be used to compute the probabilities of various seasons.
3.1 BAYES’ THEOREM
The BNs characterizes a problem domain consisting of a set of variables (attributes) E = {E1, E2, … En}. In the Bayesian terms, E is considered as the “evidence”, H is considered as the hypothesis, and data E belongs to a specific class C. For the classification problems, our goal
is to determine |, i.e. we are searching for the probability that sample E belongs to class C, given that we know the attribute and description of E.
has
disadvantages, such as the high processing time and needs training before we operate the network.
Statistical techniques are simple to implement. However, their capability is limited to specific problems. Vessel speed is a good example of a variable in which these techniques are effective because of their extreme values. In cases where anomalies are uniformly dispersed in the sample, these techniques are ineffective.
©2015: The Royal Institution of Naval Architects
A-147
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76