5.1. Experimental setup
In this section the proposed agent collaborative algorithms are evaluated in a real world WSN deployment for vehicle localization and classification. The experiments are carried out on a schoolyard. Big toy tanks
are used to simulate real vehicles. The WSN comprises 8 MICAz motes from MICAz mote developer's kit (a commercial product of Crossbow Inc.) [26
]. The MICAz is a 2.4GHz, IEEE 802.15.4 compliant module used for enabling low-power, wireless sensor networks. The mote offers sensor boards that can measure signatures such as acoustic, acceleration and so forth [26
]. In the experiments, acoustic signatures are measured by the acoustic sensors provided by the MICAz sensor board. However the mote is not equipped with sensors for seismic signatures; therefore seismic sensors are artificially connected to the mote through its sensor board interface. Moreover the MICAz motes are programmed to sample at the frequency of 1.024 kHz to accommodate acoustic and seismic signatures of high frequency.
In the experiment, the 8 MICAz motes are randomly deployed on the schoolyard as illustrated in Figure 10
. As shown in the figure, the 8 motes denoted by s1, s2 through s8 are deployed within an area approximately in the size of 26m
. Their corresponding x
coordinates in the Cartesian coordinate system are marked in the parentheses. The star in the figure denotes an imaginary target.
Following the proposed heterogeneous agent architecture, s1 through s7 are configured to be observing agents and s8 is programmed to take the role of a manager agent. Meanwhile s8 is connected to a laptop, which therefore also serves as the interface agent. In this configuration, the system can be viewed as a miniature implementation of the proposed heterogeneous agent architecture, though only 8 sensor nodes are available.
A final remark on the deployment is that s8 will not participate in measurements of any kind. It is solely responsible for coordination of the observing agents and dispatch of mobile agents.
5.2. Agent collaborative vehicle localization experiments
In the localization experiment, the vehicle is a toy tank emitting loud sound. It is kept stationary during the localization. Nevertheless it must be clarified that the proposed method is equally applicable to localization of moving vehicles. In one experiment, the vehicle is deployed at a position whose Cartesian coordinates are (-4, 3
). This is exactly where the target is in Figure 10
. The acoustic signatures measured by microphone sensors at s1, s2 and s3 (shown as examples) are shown in Figure 11
. The signatures have all been processed to eliminate the influence of sensor gain discrepancy.
Following the proposed localization algorithms, s1, s2, s3 and s5 (reporting the highest
average energy) are selected to collaborate. The average energy and the positions of these observing agents are used to determine the objective function (20)
. The steepest descent search is carried out by the manager agent s8. The search result is reported in Figure 12
. In the search, the termination condition is set to be 0.005, while the maximum search step is assigned to be 100. Numerical calculation shows no relaxation of termination condition is needed. The search starts from the position of s1 which reports the highest average energy. It takes only 10
steps to converge to the estimated location (-4.89, 2.52
). The estimate is 1.05m away from the true position. The distance between s1 and the target is 10.97m, therefore the relative localization error is no more than 1.05/10.97×100%=9.57. This result proves the proposed agent collaborative algorithm can efficiently perform vehicle localization in WSN.
As pointed out in the formulation of the agent collaborative algorithm, the selection of agents involved in the objective function (20)
is actually by intuition. To compare with the intuitive approach an experiment using the 4 agents giving the lowest average energy is conducted. The result is shown in Figure 13
. As shown in the figure, s4, s5, s6 and s7 (reporting the lowest
average energy) are chosen for collaboration and the search is started from s5 (reporting the highest average energy among the 4 agents). It takes 11 steps to converge to the estimated location (-3.07, 1.91
) which is 1.43m away from the actual position. Again no relaxation of termination condition is deed.
Note that the proposed algorithm needs 10 steps and the estimated position is 1.05m from the true location. By comparison, localization using the 4 agents with the highest average energy requires less search steps and provides better localization accuracy. Therefore the proposed search method is time efficient and accurate. Further evaluation of the proposed search method (i.e. searching with the agents reporting highest average energy) is carried out by localizing vehicles randomly positioned within the sensor field as shown in Figure 14
As shown in Figure 14
locations (denoted by T1 through T9, and marked by stars) are randomly chosen. Their corresponding estimated locations are denoted by E1 through E9 respectively (marked by diamonds). It shows that the algorithm can accurately localize vehicles at randomly selected position. Note that some of the localization results (like E1, E3, E6 and E8) are more accurate than others. This is due to differences in the relative positions between deployed sensor nodes and the vehicles.
The efficiency and accuracy of the proposed agent collaborative localization algorithm has been validated by these experiments. Its good localization accuracy also benefits the agent collaborative classification because the proposed homogeneous classification decision fusion is based upon the estimated position of the target.
5.3. Agent collaborative vehicle classification experiments
The vehicle classification experiments are performed using the same deployment in Figure 10
. The vehicles to be classified are either toy tanks or jeeps. To achieve better classification accuracy, both acoustic and seismic signatures are observed by MICAz motes. In our implementation, s1, s3, and s5 are programmed to measure acoustic signatures, while s2, s4 and s6 are configured to measure seismic signals. Once again s8 serves as the manager agent and interface agent. Note that s7 is not used. Such choice is to maintain the equilibrium between agents measuring acoustic signatures and the ones measuring seismic signatures.
Learning samples are prepared in a supervised approach. Different vehicles are placed at various positions. The seismic and acoustic signals are accordingly measured and stored by corresponding observing agents. Meanwhile an operator determines the vehicle type and instructs the manager agent to send vehicle type information to the observing agents. This ensures the locally gathered samples are correctly labeled. All sampling lasts for 2 seconds. Recall that the sampling frequency is 1.024 kHz; therefore each sample contains 2048 sample points.
When samples are prepared, the proposed distributed SVM learning can be started. Mobile agents are sent from the manager agent s8 to observing agents s1 through s6 to extract features, learn local SVM classifiers and compute hull vectors. Radial basis function (RBF) is used as the kernel for SVM learning.
Typical acoustic features of toy tanks and jeeps are demonstrated in Figure 15
. In the figure, Sn denotes the nth component of the feature vector and normalized energy is expressed in logarithm coordinate. Evidently the features make these two types of toy vehicles distinguishable, but such features are more than enough to separate toy tanks from jeeps. For example, the first component s1 is identical for both vehicles; therefore it contributes extremely little, if any at all, to the discrimination of the vehicles. Following the same principle, components s4, s5 and s6 are negligible too, because these components don't vary much for two vehicles. In this way an 8-dimensional feature vector is reduced to a 4-dimensional one. As a result it decreases the computational expense of both SVM learning and hull vectors. Furthermore the resulted support vectors and hull vectors are significantly reduced in number. Similarly the seismic features are also reduced to 4-dimensional ones.
Now that feature vectors are extracted from raw data and artificially reduced, they can be used to learn local SVM classifiers and calculate corresponding hull vectors. The samples at each observing agent are split into two parts, one for SVM training and the other for testing. The sample splitting and performance of learned SVM are listed in Table 1
. In Table 2
, the statistics concerning derived support vectors and hull vectors are reported.
In Table 1
, ‘T’ denotes a toy tank and ‘J’ refers to a toy jeep. ‘Training#’ and ‘Testing#’ are the numbers of samples used to train and test the SVM classifiers respectively. Accuracy shows how well the learned SVM classifies the testing samples. Obviously SVM classifiers learned from different samples (even if they are of the same modality) present contrasting accuracy. This is not unusual, because as said above, each observing agent can only observe a portion of the whole characteristics of a target. That is also why agent collaboration is needed to enhance classification accuracy.
In table 2
, ‘SV#’ denotes the number of support vectors (SV), ‘HV #’ means the number of hull vectors (HV). ‘HSV #’ refers to the number of the vectors in the union of support vectors and hull vectors. We are most interested in the difference between SV# and HSV#. Because SV# indirectly represents the transmitted data volume for the SV only algorithm and HSV #corresponds to that of the HV and SV algorithm. In all cases, HSV# is larger than SV#, which is expected by theoretical analysis. Accordingly the HV and SV algorithm consumes more energy for communication between the manager and observing agents. As noted before, hull vectors are incorporated to improve distributed learning accuracy. Now let us turn to see its learning performance as shown in Table 3
In the table, ‘Centralized’ refers to the traditional centralized learning algorithm that sends all samples to a central point and learns the global SVM classifier accordingly. Evidently the centralized algorithm performs the best for both modalities, because it has access to all available samples. The HV and SV algorithm is almost as good as the centralized algorithm. In contrast, the SV only algorithm performs much worse than the centralized. For acoustic features, it is nearly 7% worse than the centralized. For seismic, it is almost 10% worse in performance.
Therefore as far as learning accuracy is concerned, the HV and SV algorithm proposed in this paper is superior to the SV only algorithm. However just as shown in Table 2
, the accuracy improvement is at the cost of increase in communication load. Usually such tradeoff is desirable, because in classification applications, accuracy is of the highest priority compared to other factors like energy consumed by wireless communication.
Furthermore it should be noted that data transmission has been drastically decreased by mobile agents. Feature extraction and the HV and SV algorithm both perform data compression. Take transmission of raw data of s1 for example. All together, 624kByte raw data have to be transmitted (note that, a double point takes 4 bytes; each raw sample is 2048*4 bytes; there are 40+38 samples all together). However, for the HV and SV algorithm using mobile agents, only 0.248kByte data (volume of the derived 4-dimensional hull vectors and support vectors) need to be transmitted. Even if dispatch of mobile agents is taken into account, the data transmission is still much smaller than 624kBytes.
Now that global SVM classifiers have been learned at the manager agent using acoustic and seismic features respectively, they can be used to classify unknown targets detected in the sensor field.
The classification of an unknown target is relatively easier than the learning of the classifier itself. When a target is detected and located, the manager agent instructs the observing agents to measure either acoustic or seismic signatures. Then mobile agents for feature extraction are sent to these observing agents. The extracted features are sent back to the manager agent where homogeneous and heterogeneous fusions are carried out to make the fused classification decision.
In the experiment, s1, s3 and s5 measure acoustic signature; s2, s4 and s6 observe seismic signatures. A toy tank is place at position (-4,-5)
whose estimated position is (-5.3, -4.3)
. The homogeneous and heterogeneous fusion results are presented in Table 4
following the proposed fusion algorithms.
In the table, the weight for homogenous fusion is determined following (23)
but normalized. The modality weight for heterogeneous fusion is determined upon their classification performance. In this experiment, as shown in Table 3
, the classification accuracy is 93.59% (using the HV and SV algorithm) for acoustic modality and 83.33% for seismic modality. Therefore the normalized modality weight for acoustic classification is 93.59% / (93.59%+83.33%) =0.529. The seismic modality weight is similarly calculated.
According to the decision rule, a positive decision means a toy tank while a negative one means a toy jeep. For acoustic classification, s1 and s3 believe the target is a toy tank, but s5 reports it as a toy jeep. Disputes arise and fusion needs to be made to form a more reliable decision. The situation is similar for seismic classification. Homogeneous fusion results are further fused following (24)
and the final decision is 1.0638. It means the global fusion decision is a toy tank. Such decision is in accordance with the truth, for it is known a priori that the target is a toy tank.
The fusion results in Table 4
show that local decision may be incorrect, but following the proposed fusion algorithm a reliable global decision is obtained. Therefore the proposed method is efficient to perform hierarchical fusion of both homogenous and heterogeneous decisions.
As corroborated by the experiment results, the proposed heterogeneous agent architecture significantly facilitates designs of WSN and remarkably reduces in-network communication load. With this architecture, the proposed localization and classification algorithms are easily implemented and prove to be accurate and energy efficient.