4.3. Results
To conduct our experiment, we created a realistic scenario comprising an ESP32 client and a gateway equipped with SEMTECH SX1276 and SX1278 modules implementing a LoRa physical layer. We established communication between the client and gateway nodes under 433 MHz and 915 MHz frequencies (as described in [
14]) using two different antennas. We considered 10, 11, and 12 (we ignored 7, 8, and 9 SF values due to previous measurements and the correlation between lower SF values had implied in worse transmitter sensitivities [
30]) SF values in this evaluation. Concerning bandwidth (BW), we considered the following values defined in LoRaWAN specification. We decided to use a fixed value of
ratio for the coding rate (CR) throughout all experiments in order to simplify the execution of the tests.
Figure 6 illustrates the evaluation scenario involving communication from the Computer Science Department to the CRITT Innovation Center on the Federal University of Juiz de Fora campus. The obtained results are summarized in
Table 3. Since there are correlations between SNR values for distances from 1 and 800 m, we can conclude that in the evaluated outdoor environment, measures of approximately 10 m of distance between the nodes are enough to demonstrate the need for node-aware parameter changes to improve SNR. The graphs from
Figure 11,
Figure 12,
Figure 13 and
Figure 14 illustrate the change/shift points, and the biggest three SNR improvements were highlighted with an arrow in each figure (We only present a sample of the measures at this moment. We complement these results later using tables).
The graphs introduced by this section show a comparison of both techniques: without any windows, referring to the technique by [
14] (
Figure 11), and with windows of 10 and 20 units of length (
Figure 12 and
Figure 13), using
and
.
As illustrated in
Figure 11,
Figure 12,
Figure 13 and
Figure 14, we noted that for the presented configuration, there is a direct correlation between the maximum SNR values and the window size. As the length of the window becomes larger, there are slightly higher maximum SNR values measured, especially considering the 20 elements’ window.
Considering the previous related work, InstantChange and SlidingChange algorithms were compared with LR-ADR to complement and enrich our analysis. All the algorithms were implemented to conduct the comparison. Despite LR-ADR being developed to select the best network server in a LoRa network with multiple gateways, we compare our approach with it, considering the gain in terms of SNR. The LR-ADR uses simple linear regression to smooth the SNR signal from the last packets received to introduce the LR-ADR. The authors consider, by default, 10 SNR values for each gateway per end device. In our case study, in a real testbed with one gateway and real data collected, we consider different SF, BW, and Frequency values as a base to apply the LR-ADR and configure it. The configuration shifts and SNR metrics were computed and compared with our proposal.
Considering LR-ADR algorithm, the
is calculated in time
T as follows:
where
is the
at time
T, and
is equal to
. The
is represented by:
To estimate the next expected SNR values and choose the parameters’ values, the authors use
More details about the LR-ADR algorithms can be found in [
28].
Table 3 shows the best result for each parameter value combination in the experiment at a specific moment, and
Table 4 shows the number of shifts of each setting to another.
Table 3 shows all observed configurations and their corresponding SNR gains. The average from all 27 measurements taken on the given setup (i.e., windowless, a window of 10 units of length, etc.) is displayed in the “Average” column. Measurements 1, 2, and 3 show the first, second, and third biggest SNR gain variations from each setup with a sliding window, respectively. The highlighted values show the highest SNR gain (in this case, 11.89%) and the best gain average (of 4.60%) observed.
From
Table 3, it is possible to note a slight rise in the SNR gain for the technique with a window using
and a smaller rise for the window with
. The window with
presented a significantly smaller gain compared with the previous two window lengths. Then, the window with
scored the smallest gain of all techniques. Observing the data on
Table 3, it is possible to note that the window with
scored the best results from the entire table, reaching an SNR gain of 11.89%. However, as the average values show, the window with
achieved the best gain consistency. Finally, we can highlight that LR-ADR obtained a result close to not using windows. Considering that the LR-ADR is based on simple linear regression with smoothing for the last SNR measure, the results are very intuitive, keeping the tendency of the next SNR value, which is not always the best configuration.
Table 4 shows the frequency of the shifts that occurred and the average of nine measurements from the changes for each setup. The highlighted values represent the biggest and smallest frequency shift averages. From an analysis of
Table 4, it is possible to note that the SlidingChange technique with the window length equal to 10 elements, considering the average values, reached the highest number of carrier changes, scoring even higher values than the InstantChange technique. The technique with
elements had a moderate reduction, while the windows with 30 and 40 elements presented the most significant changes compared to the InstantChange technique. It is worth noting that the technique with 40 elements scored a reduction ratio of more than
on carrier changes compared with the InstantChange technique. The LR-ADR obtained the worst result, generating the highest number of configuration shifts, even more than the configuration with window
. We believe that this is due to the use of the trend of the value of the next SNR found by the linear regression method simple, which is not always correct in environments with mobility, as is the case of the evaluated environment.
Based on the data from
Table 3 and
Table 4,
Table 5 illustrates the increase in SNR gains by percentage when compared with the results obtained from the InstantChange technique presented in [
14]. For each window length evaluated in this paper, it showed a reduction in the number of changes in the carrier modulation compared with the InstantChange technique. From the analysis of
Table 5, we observed that the window with
is the one that achieved the best SNR gains, scoring an improvement of 44.89% compared with the InstantChange technique. However, it reached a higher number of frequency changes, as 37.80% more changes were recorded regarding that parameter.
In contrast, compared to the InstantChange technique, the sliding window of is the one that achieved the highest reduction on the average number of carrier changes, scoring 53.25% fewer changes. However, the sliding change technique showed the smallest SNR gain, scoring a reduction of 8.92% of the average SNR gain.
The window with
was represented by the InstantChange technique. From the analysis of the results, we concluded that the larger sliding window does not always mean a higher SNR gain. We identified the most efficient setups have between 0 and 30 units of length. The sliding window of 20 units of length could be considered the one with the best efficiency ratio since it achieved a considerable reduction in the carrier frequency changes and still had an SNR gain. Considering the LR-ADR, we would like to highlight that the proposed technique by [
28] was created to be applied in a LoRa network with more than one gateway. Once it is not our scenario, the LR-ADR was evaluated, keeping the parameters (i.e., SF, BW, Frequency) values or shifting to another one. This particularity of our evaluation and proposal objectives could have influenced the LR-ADR performance’s final result.