1:
2: % Statistical Analysis of the collected data through measurements
3:
4: This section presents the statistical analysis of the data collected
5: by campaign described in Section \ref{S:Experiment}. We first
6: estimate the path loss model parameters of (\ref{E:meanPathLoss})
7: and (\ref{E:logNormalShadowing}). Next, we analyze the shadowing
8: loss correlations which exist on different pairs of links.
9:
10: %----------------------------------------------------------------------
11:
12: \subsection{Analysis of Received Power}\label{S:Analysis_RSS}
13:
14: We denote the number of the deployment experiment as $m\in
15: \{1,\ldots,M\}$, where $M$ is the number of deployments (here,
16: $M=15$). We denote the set of frequencies measured as
17: $\mathfrak{F}$. The received signal power between node $i$ and node
18: $j$ for experiment $m$ at center frequency $f\in \mathfrak{F}$ is
19: denoted $P_{i,j}^{(m)}(f)$ and can be written using
20: (\ref{E:logNormalShadowing}) and (\ref{E:shadowFadingDefn}) as
21: \begin{equation}\label{E:freqDependentRSS}
22: P_{i,j}^{(m)}(f) = P_{T_j} - \Pi_0 - 10n_p\log \frac{d_{i,j}}{\Delta_0} -
23: X_{i,j}^{(m)} - Y_{i,j}^{(m)}(f),
24: \end{equation}
25: where $Y_{i,j}^{(m)}(f)$ is the non-shadow fading and
26: $X_{i,j}^{(m)}$ is the shadow fading on link $(i,j)$ during
27: experiment $m$. Shadow fading is considered to be constant across
28: the frequency band, as discussed in Section \ref{S:Introduction}. We
29: denote the frequency average received power as $P_{i,j}^{(m)}$,
30: \begin{equation*}
31: P_{i,j}^{(m)} \triangleq \frac{1}{|\mathfrak{F}|}\sum_{f\in \mathfrak{F}}
32: P_{i,j}^{(m)}(f).
33: \end{equation*}
34: From (\ref{E:freqDependentRSS}), we can write $P_{i,j}^{(m)}$ as,
35: \begin{equation}\label{E:ApproxRSS}
36: P_{i,j}^{(m)} = P_{T_j} - \Pi_0 - 10n_p\log \frac{d_{i,j}}{\Delta_0} -
37: X_{i,j}^{(m)} - \frac{1}{|\mathfrak{F}|}\sum_{f\in \mathfrak{F}}
38: Y_{i,j}^{(m)}(f).
39: \end{equation}
40: In other words, (\ref{E:ApproxRSS}) can be written as,
41: \begin{equation}\label{E:link_receivedSignal_eachExp}
42: P_{i,j}^{(m)} = P_{T_j} - \Pi_0 - 10n_p\log \frac{d_{i,j}}{\Delta_0} -
43: X_{i,j}^{(m)} - Y_{i,j}^{(m)},
44: \end{equation}
45: where $Y_{i,j}^{(m)} = \frac{1}{|\mathfrak{F}|}\sum_{f\in
46: \mathfrak{F}} Y_{i,j}^{(m)}(f)$. Because $Y_{i,j}^{(m)}$ is an
47: average of measurements at many different frequencies, we argue that
48: it may be well-represented as Gaussian (in dB), regardless of the
49: underlying frequency-selective fading mechanism (\eg, Rayleigh or
50: Rician). Since $X_{i,j}^{(m)}$ is also log-normal \cite{coulson}, we expect the sum $Z_{i,j}^{(m)}$ to
51: also be Gaussian (in dB).
52:
53: A linear regression of the frequency averaged received signal powers
54: $\{P_{i,j}^{(m)}\}_{i,j}$ versus known distances $\{ d_{i,j}
55: \}_{i,j}$ is used to estimate the constants $(P_{T} - \Pi_0)$ and
56: $n_p$ (\ref{E:link_receivedSignal_eachExp}) for each experiment $m$.
57: In our experiments, we have used $\Delta_0 = 1$m. Since all nodes
58: are set to the same transmit power and have approximately equal
59: battery voltages, and since we estimate $(P_{T} - \Pi_0)$ in
60: addition to $n_p$, we are not required to know the exact transmit
61: power $P_T$ at the current battery voltage of the nodes in the
62: network during experiment $m$. The linear regression also determines
63: the variance of $Z_{i,j}^{(m)}$.
64:
65: %-----------------------------------------------------------------------
66:
67: \subsection{Analysis of Link Correlations}
68: \label{S:Analysis_corr}
69:
70: In this subsection, we describe the computation of the correlation in
71: fading between pairs of links. This requires computing correlation
72: in the sample values of $Z_{i,j}^{(m)}$ for different pairs of links
73: $(i,j)$ as described in Section \ref{S:Analysis_RSS}.
74:
75: \subsubsection{Similar Geometry Links}
76: We use the term ``link geometry'' to describe for two links, link
77: $a$ and link $b$, the relative coordinates of the end points of the
78: two links. In a grid network, there can be many pairs of links with
79: the same link geometry (within a rotation). As one example, the link
80: pair of link $a$ and link $b$ shown in Fig.~\ref{F:Examples}, is
81: repeated 16 times in the network as shown in
82: Fig.~\ref{F:SimilarGeoLink}.
83:
84: \showfigure{
85: \begin{figure}[htbp]
86: % Requires \usepackage{graphicx}
87: \centerline{
88: \psfig{figure=similarGeometryConstruction.eps,width=1.4in}}
89: \caption{Link pairs with identical link geometry in a grid deployment. Each link pair
90: is shown with one link as a dotted (-~-) line and another link as a
91: solid lines (--). All link pairs with identical link geometry are
92: shown.}
93: \label{F:SimilarGeoLink}
94: \end{figure}
95: }
96:
97: Let $L$ denote the number of times a particular link geometry is
98: repeated in the network. We denote the $p^{\mbox{th}}$ link pair as
99: the two links $(i_p,j_p)$ and $(k_p,l_p)$, where $p\in
100: \{1,\ldots,L\}$. Then $Z_{i_p,j_p}^{(m)}$ and $Z_{k_p,l_p}^{(m)}$,
101: where $m\in \{1,\ldots, M\}$, represent the total fading on the
102: $p^{\mbox{th}}$ repeated link pair for experiment number $m$. Then
103: vectors $\mbZ_a^{(m)}$ and $\mbZ_b^{(m)}$ are defined as
104: \begin{equation}\label{E:Z_ab^mDefn}
105: \mbZ_a^{(m)} =
106: [Z_{i_1,j_1}^{(m)}, \ldots ,Z_{i_L,j_L}^{(m)}]^T \quad,\quad
107: \mbZ_b^{(m)} = [Z_{k_1,l_1}^{(m)}, \ldots
108: ,Z_{k_L,l_L}^{(m)}]^T.
109: \end{equation}
110: We then define vectors $\mbZ_{a}$ and $\mbZ_{b}$ as
111: \begin{equation}\label{E:Z_abDefn}
112: \mbZ_{a} =
113: [\mbZ_{a}^{(1)^T},\ldots,\mbZ_{a}^{(M)^T}]^T \quad,\quad
114: \mbZ_{b} =
115: [\mbZ_{b}^{(1)^T},\ldots,\mbZ_{b}^{(M)^T}]^T.
116: \end{equation}
117: Vectors $\mbZ_{a}$ and $\mbZ_{b}$ are both $LM$x1 sized vectors.
118: Together they contain all measured total fading values for pairs of
119: links which share a particular link geometry. The correlation
120: coefficient of total fading on link $a$ and link $b$,
121: $\rho_{Z_a,Z_b}$, can be computed by taking vectors $\mbZ_{a}$ and
122: $\mbZ_{b}$ as sample values of total fading for link $a$ and link
123: $b$ respectively.
124:
125: We have computed the correlation coefficient for total fading on
126: link $a$ and link $b$ for a variety of link geometries.
127: Table~\ref{T:correlationComparision} shows the results for various
128: link pair geometries. We also run a hypothesis test to determine if
129: the measured correlation is statistically significant. This test
130: compares hypotheses,
131: \begin{eqnarray*}
132: &H_0:& \mbox{$Z_a$ and $Z_b$ have $\rho=0$},\\
133: &H_1:& \mbox{$Z_a$ and $Z_b$ have $\rho\neq0$}.
134: \end{eqnarray*}
135: We report $\PR{\mbox{measuring }\rho|H_0}$ using the method
136: described in \cite[pp. 427-431]{Hines4th2003}, in
137: Table~\ref{T:correlationComparision}. The proposed correlated link
138: shadowing model and the Gudmundson model, also mentioned in
139: Table~\ref{T:correlationComparision}, will be discussed in Section
140: \ref{S:CorrelationModelDes}.
141:
142: \begin{table}[p]
143: \begin{center}
144: \begin{tabular}{|c|c|l|c|c||c|c|l|c|c|}
145: \hline
146: % after \\: \hline or \cline{col1-col2} \cline{col3-col4} ...
147: & Geometry & \multicolumn{3}{c|}{Correlation $\rho$} & & Geometry & \multicolumn{3}{c|}{Correlation $\rho$}\\
148: \hline
149: & & Meas- & Prop. & Gud. & & & Meas- & Prop. & Gud.\\
150: & & ured & Model & model & & & ured & Model & model \\
151: \hline
152: 1 & \psfig{figure=geometry1.eps,width=0.5in} &0.33*** &0.21 & 0.13 & 15& \psfig{figure=geometry15.eps,width=0.5in} &-0.04 &0.05& 0.04 \\
153: \hline
154: 2 & \psfig{figure=geometry2.eps,width=0.5in} &0.21*** &0.17 & 0.04 & 16& \psfig{figure=geometry16.eps,width=0.5in} &0.12*** &0.10& 0.08 \\
155: \hline
156: 3 & \psfig{figure=geometry3.eps,width=0.5in} &0.23*** &0.24 & 0.13 & 17& \psfig{figure=geometry17.eps,width=0.5in} &0.08* &0.07& 0.08 \\
157: \hline
158: 4 & \psfig{figure=geometry4.eps,width=0.5in} &0.05 &0.03 & 0.04 & 18& \psfig{figure=geometry18.eps,width=0.5in} &0.12*** &0.11& 0.04 \\
159: \hline
160: 5 & \psfig{figure=geometry5.eps,width=0.5in} &0.17*** &0.19 & n/a & 19& \psfig{figure=geometry19.eps,width=0.5in} &0.03 &0.10& 0.08 \\
161: \hline
162: 6 & \psfig{figure=geometry6.eps,width=0.5in} &-0.05 &0.00 & n/a & 20& \psfig{figure=geometry20.eps,width=0.5in} &0.21*** &0.13& 0.13 \\
163: \hline
164: 7 & \psfig{figure=geometry7.eps,width=0.5in} &-0.01 &0.00 & n/a & 21& \psfig{figure=geometry21.eps,width=0.5in} &-0.02 &0.08& 0.04 \\
165: \hline
166: 8 & \psfig{figure=geometry8.eps,width=0.5in} &-0.10** &0.00 & n/a & 22& \psfig{figure=geometry22.eps,width=0.5in} &0.23*** &0.16& 0.13 \\
167: \hline
168: 9 & \psfig{figure=geometry9.eps,width=0.5in} &-0.03 &0.05 & 0.04& 23& \psfig{figure=geometry23.eps,width=0.5in} &0.00 &0.05& 0.04 \\
169: \hline
170: 10&\psfig{figure=geometry10.eps,width=0.5in}&0.18*** &0.21 & 0.08& 24& \psfig{figure=geometry24.eps,width=0.5in} &0.06 &0.16& 0.08 \\
171: \hline
172: 11& \psfig{figure=geometry11.eps,width=0.5in}&0.04* &0.08 & 0.13& 25& \psfig{figure=geometry25.eps,width=0.5in} &0.08** &0.13 & n/a \\
173: \hline
174: 12& \psfig{figure=geometry12.eps,width=0.5in} &0.14***&0.08 & 0.13& 26& \psfig{figure=geometry26.eps,width=0.5in} &0.12 &0.16 & n/a \\
175: \hline
176: 13& \psfig{figure=geometry13.eps,width=0.5in} &0.17***&0.08 & 0.13& 27& \psfig{figure=geometry27.eps,width=0.5in} &0.08 &0.00 & n/a \\
177: \hline
178: 14& \psfig{figure=geometry14.eps,width=0.5in} &0.05 &0.06 & 0.08& 28& \psfig{figure=geometry28.eps,width=0.5in} &0.03 &0.02 & 0.02 \\
179: \hline
180: \end{tabular}
181: \\
182: \begin{tabular}{|c c c|}
183: \hline \multicolumn{3}{|c|}{p- value or $\PR{\text{getting measured
184: $\rho | H_0$}}$}\\
185: \hline
186: *** $p < 0.005$ & $\quad$ ** $p < 0.01$ & $\quad$ * $p < 0.05$\\
187: \hline
188: \end{tabular}
189: \end{center}
190: \caption{Link Geometry and Correlation Coefficients (Observed, Proposed Model, and Model of [Gudmundson 1991])}
191: \label{T:correlationComparision}
192: \end{table}
193:
194:
195: \subsection{Discussion}
196: The results show that, for many link pair geometries, it is
197: extremely unlikely that the fading losses measured on the pair of
198: links are independent. For 15 of the 28 studied link geometries,
199: there is statistically significant non-zero correlation. Those 15
200: links are consistently those geometries in which the two links are
201: proximate, \ie, their lines from transmitter to receiver partially
202: overlap, or nearly overlap. The likelihood that the measured
203: correlation coefficient was measured by chance in the case when
204: $\rho=0$ is extremely small, \ie, less than 0.5\%, for 11 of the 15
205: link geometries which showed correlation.
206:
207: Also note that the correlation coefficients are relatively large in
208: magnitude. The highest $\rho$ is 0.33, six link geometries have
209: $\rho > 0.20$, and eleven link geometries have $\rho > 0.10$. Fading
210: loss on one link is obviously not purely determined by the losses
211: experienced on its geographically proximate links; however, the
212: correlation coefficient indicates that knowing the losses on the
213: proximate links can give quite a bit of information about the loss
214: on that one link.
215: