VI. JUSTIFICACION
3. CAPÍTULO III: DISEÑO DEL SISTEMA
3.1. DISEÑO ARQUITECTONICO
3.1.1. Diseño arquitectónico basado en el modelo de vista 4+1 de Krutchen
3.1.1.1. Descripción de la arquitectura en contexto
Flickering LED Lights Stalled motor fan Repetitive noise
Unknown increase in variance
This subsection will introduce the anomalies and apply the various techniques in the anomaly warning and fault detection. The results were evaluated and discussed.
Detection of Flickering LED Lights
One of the common degeneration issues with lights is that the lights will start to flicker. Below showed the waveform of a 1-second block of flickering LED lights and the extracted alternate between on and off frequently due to the flickering instead of maintaining stable constant current.
Mean Variance Range Largest Gradient
0.1428 0.0223 0.48 0.32
119
The features of this 1-second block formed the test object for the kNN algorithm. It was injected into the environment to be classify and test for faults and anomalies using the proposed methods in this paper. All the 4 features were used in the algorithm, however, due to the limitation of representation,
Figure 4.57 will only show the 3-D environment using mean, variance and range as the axis.
120
TABLE 4.24EVALUATION OF TEST DATA OF FLICKERING LEDLIGHTS
Technique 1: Based on percentage on score of prediction
Highest score Required score Anomaly
warning?
69.23% (Class 7) 75% YES
Technique 2: Based on distance from centroid Distance of test
object to centroid Maximum distance
Anomaly warning?
0.3347 0.0625 YES
Technique 3: Based on average distance from k nearest neighbours
Average distance of test object with
neighbours
Maximum
distance Anomaly warning?
1.3070 0.2804 YES
Although the correct class for LED lights was class 4 but the kNN algorithm had incorrectly predicted the test object to be labelled as class 7 with a score of 69.23%; class 4 had a score of 30.77%. This was because the features extracted from the flickering LED lights were no longer representative of a normal working LED light. This percentage was less than the required score of 75% to be classified as normal operation thus it did not pass the criterion and an anomaly warning was issued.
In the second test, the distance of the test object from the centroid exceeded the acceptable boundary which was the average distance between the centroid and the elements in the cluster plus 3 times its standard deviation. Thus, an anomaly warning was also issued for this test technique.
The 3rd and last technique used was to compare the average distance of the test object
with its k nearest neighbours in the labelled cluster to the average distance between the elements with their k nearest neighbours. Anomaly warning was also triggered as it exceeded the maximum distance.
Detection of Stalling Motor Fan
A stalling motor can be caused by an obstruction of an unknown object on the blades. This abnormal operation was simulated by blocking the blades of the DC motor fan. Figure 4.58 shows the captured waveform of the motor stalling while hindered by the obstruction.
121
Mean Variance Range Largest Gradient
0.3457 0.0012 0.4230 0.2576
FIGURE 4.58CURRENT WAVEFORM OF STALLED DCMOTOR FAN
The test object representing the stalling DC motor fan was described by the 4 extracted features. Figure 4.59 shows the test object in the test environment where kNN algorithm and the proposed 3 techniques for anomaly warning and fault detection were applied.
FIGURE 4.59TEST OBJECT OF STALLED DCMOTOR FAN
The kNN classification algorithm labelled this test object as Class 8, which was incorrected as DC motor fan was Class 5. It did not matter as the features were not representative of it as a normal operating DC motor fan.
The test object representing the stalling DC motor fan did not meet the requirement of the 3 anomaly criteria. Anomaly warnings were issued for all techniques.
122
TABLE 4.25 EVALUATION OF TEST DATA OF STALLED FAN
Technique 1: Based on percentage on score of prediction
Highest score Required score Anomaly
warning?
69.23% (Class 8) 75% YES
Technique 2: Based on distance from centroid Distance of test
object to centroid Maximum distance
Anomaly warning?
0.2192 0.0392 YES
Technique 3: Based on average distance from k nearest neighbours
Average distance of test object with
neighbours
Maximum
distance Anomaly warning?
0.2995 0.2723 YES
Detection of Repetitive Noise
Another common anomaly in a branch circuit is the occurrence of repetitive noise in the
normal operation. The waveform is shown in Figure 4.60.
Mean Variance Range Largest
Gradient
0.3247 0.0012 0.32 0.24
FIGURE 4.60CURRENT WAVEFORM WITH REPETITIVE NOISE
The test object, created from the 4 extracted features in the 1-second block, representing the noisy waveform, was injected in the test environment as shown in Figure 4.61 below.
123
FIGURE 4.61TEST OBJECT OF REPETITIVE NOISE
TABLE 4.26EVALUATION OF TEST DATA OF REPETITIVE NOISE ON LEDLIGHTS WAVEFORM
Technique 1: Based on percentage on score of prediction
Highest score Required score Anomaly
warning?
52.85% (Class 8) 75% YES
Technique 2: Based on distance from centroid Distance of test
object to centroid Maximum distance
Anomaly warning?
0.2493 0.0392 YES
Technique 3: Based on average distance from k nearest neighbours
Average distance of test object with
neighbours
Maximum
distance Anomaly warning?
0.2866 0.2723 YES
The kNN classification algorithm incorrect labelled this as Class 5 instead of Class 4 LED lights. The score was only 42.86% which was much lower than the required 75%. The distance of the test object to the centroid also exceeded the boundary and the average distance of the test object with its k nearest neighbour was also larger than the maximum distance. Therefore, anomaly warnings were triggered for all techniques.
Detection of Unknown Large Variance
A tricky scenario with anomaly is the change in variance while maintaining the same mean value of the current waveform. This could be due to degeneration of the loads. In this situation, the common meter, which can only measure the mean value, will not be able to identify any anomaly as the mean value of the grid does not change. An example is as seen in
124
Figure 4.62 below. The larger variance occurred in Class 7 which had both Humidifier and Fan switched on.
Mean Variance Range Largest Gradient
0.3247 0.0013 0.32 0.24
FIGURE 4.62CURRENT WAVEFORM WITH LARGE UNKNOWN VARIANCE
The test object, described by the four extracted features in the 1-second block, representing the waveform with the larger variance was injected in the test environment as shown below.
FIGURE 4.63TEST OBJECT OF LARGE VARIANCE
It was observed that even though the mean value of the test object falls along the same line as the mean of Class 7, but the difference in variance pulled the test object away from the cluster. Technique 1, which uses the percentage on score of prediction indicated 100% and no
125
anomaly warning. However, as the test object was pulled very far away from the centroids and its neighbours, it failed in the other two anomaly warning and fault detection techniques.
TABLE 4.27EVALUATION OF TEST DATA OF LARGE VARIANCE ON LEDLIGHTS WAVEFORM
Technique 1: Based on percentage on score of prediction
Highest score Required score Anomaly
warning?
100% (Class 7) 75% No
Technique 2: Based on distance from centroid Distance of test
object to centroid Maximum distance
Anomaly warning?
0.1541 0.0625 YES
Technique 3: Based on average distance from k nearest neighbours
Average distance of test object with
neighbours
Maximum
distance Anomaly warning?
1.0849 0.2804 YES
Summary of Anomaly Warning with Enhanced kNN Technique
This research combines three additional anomaly criteria on top of the kNN classification algorithm for a single sensor multiple load DC pico-grid. This enhanced kNN technique are put to test in a DC pico-grid containing humidifier, LED lights and fan. They are based on the percentage of the labelled cluster, based on the distance of test object from the
This research extracted 4 features (mean, variance, range and largest gradient) from the 1-
environment. The enhanced kNN technique is mostly successful in triggering warnings in abnormal operations of flickering LED lights, stalling DC motor fan, repetitive noise and larger than usual variance waveforms. These added capabilities of anomaly warning and fault detection of the kNN algorithm can be used in conjunction with load classification and monitoring of the power system in a branch circuit. This technique in this research can be valuable in predictive maintenance and early fault detection.
In the next section, this technique is further improved to a Hierarchical Enhanced k- Nearest Neighbours technique and applied to a remote monitoring system.
126
4.8 Remote Load Classification and Anomalies Warning using Hierarchical Enhanced k-