Exam Dec 2021 Flashcards
For a signal with the frequency f = 100 Hz, suggest a reasonable sampling frequency according to shannon theorem.
According to Shannon theorem, we need at least 2f. In order to have a good reconstruction, it is recommended 10f. Fs= 1000 Hz
A signal of 100 Hz is recorded during 5 seconds with a sampling frequency of 8000 Hz. What is the total number of samples obtained? [5p]
5 x 8000 = 40 00
What are the consequences of not obeying the Shannon sampling criterion? [5p]
A too low frequency creates aliasing and distortions in the reconstructed signal. A too high sampling frequency results in a good
quality of reconstruction but excessive storage,
Imagine we want to build a controller for smart curtains in a smart home. Which sensor are you going to use? For this sensor, you will have
to perform first a calibration.
Please describe what calibration exactly means. [5p]
passive opto sensor, placed outside, for measuring the ambient light.
unknown in the beginning at what sensor values the curtains should be closed and opened -> perform a calibration aka -> mapping sensor values to actual brightness levels/times of day:
- by observing the sensor values
throughout the day and map them in a table to the current light conditions (e.g., dawn, sunrise, …) and/or time of day.
appropriate values can be read from the table + used as a threshold in the controller.
better system that works reliably in different weather conditions -> this mapping of the sensor values could be
performed multiple times on different days, in order to get an average value for each light condition.
The goal of an Maneuvering Characteristics Augmentation System (MCAS) is to maintain an airplane in an horizontal position and decrease
the pitch-up tendency . Basically it measures the Angle-of-attack (AoA) and if this is larger than 14 degrees, the system activates some
stabilizer motors in the tail, to push down the aircraft nose.
Below you can see the feedback control diagram for this system. Fill in the correct
names of its components, labelled with 1 to 6. [5p]
1 is the goal, the desired value. In this case this would be 14 degrees (the desired angle of attack)
2 is the comparator, comparing this value to the value coming from 6 and sending the resulting error to 3
3 is the controller, this component gets the value from 2, and decides on that what action 4 should perform (and whether it
should perform an action at all)
4 is the actuator, in this case the stabilizer motors in the tail
5 is the output of the controlled variable, which is here the Angle-of-Attack. It is called controlled variable because the system,
or more specifically the actuator 4, has an influence on this variable
6 is the sensor, it measures the Angle-of-Attack and feeds this value to the comparator. As this sensor is observing the variable
that is controlled by the system (4), this is a closed-loop feedback control system
We record a heart activity (ECG) signal and we notice that it has some noise. A spectral analysis using FFT shows that the frequency of the
noise is 50 Hz. What kind of filter do you suggest to eliminate this noise?
A notch filter
Imagine we want optically recognize handwritten characters. Can you suggest a method to solve this problem?
Enumerate the steps you
need to take, specify which kind of classifier are you going to use and which features.
What are the advantages and disadvantages of your
approach ? [10p]
For my approach, the first step is to preprocess the images of the individial characters in order to remove noise and convert the
color or grayscale image into a binary image, where all the background is 0 and only the pixels of the actual character are 1.
This could be done by binarizing with a certain threshold, and then performing opening to get rid of any small noise
As a classifier, I would use a neural network (machine learning). In order for this to work, I would need two things:
1: features of the images that are used as inputs
2: a training set of images with the known targets (characters)
As features, I would not simply use the value of each of the pixels as input but rather statistical image moments, more
specifically the 7 invariant moments of hu. This would make the resulting classifier a lot more versatile, as new images could
be in different resolutions or written by someone else and still be correctly classified (at least with a good neural network and a
training set that is versatile enough).
The neural network then would have 7 input nodes; one for each of the 7 invariant moments of hu.
In order to classify both uppercase and lowercase letters, as well as numbers, the neural network would need 62 output
neurons. Between those input and output neurons are some hidden layers. The number of these layers would depend on the
available processing power of the system.
This neural network then would need to be trained with a set of preprocessed images showing handwritten characters (or more
specifically the invariant moments of hu of these images) as well as the targets, so the character they actually show. This set
needs to be very large in order to get a classifier with a good accuracy.
When complete, this network could be given an image of a handwritten character, and the resulting output would classify the
actual character shown. (all the output nodes can only ouput 0 or 1, so the output for a lowercase ‘a’ could be 10000000000…,
lowercase ‘b’ could be 01000000000…, and so on)
The main advantage of using a neural network as a classifier is a (possibly) very good accuracy. However, it needs a lot of
training images in order to be accurate, and requires a lot of processing power for the training.
What are Hidden Markov Models (HMMs) and where can we use them in pervasive computing? [5p]
HMMs are graphical probabilistic reasoning algorithms that can be used for example for activity classification. They contain hidden
states and observable states connected with edges. The challenge is to infer the hidden states knowing the visible observable
states
We perform an experiment for pedestrian recognition on the road. We had 1000 training images, 900 with pedestrians and 100 with nonpedestrians. For each of the following components in a confusion matrix, True Positives (TP), True Negatives (TN), False Positives (FP) and False Negatives (FN), give some values that characterize a successful classification.
When implemented in a self driving car, the True Positive Rate (which can be calculated by dividing the True Positives by 900
(the amount of actual positives, i.e. images with pedestrians)) needs to be as high as possible, as it denotes the accuracy for
correctly detecting a pedestrian. Subsequently, the false negative rate needs to be as low as possible (as it is the resulting
“rest” that did not get detected, and thereby possibly injured from the car)
The False Positive Rate (which is the amount of False Positives divided by the amount of images with no pedestrians) is not
as important in this case. It still should not be overlooked, as a False Positive would mean that the car stops (maybe even
abrubtly) on the road, which could also be dangerous.
In general, a good classification would mean a number of True Positives that is close to 1000, and a number of True Negatives
that is close to 100. Therefore, the amount of False Positives and False Negatives should be as close to 0 as possible.
When this pedestrian recognition system is used in a fully autonomous car then I would say that the True Positives should be
900 (therefore the False Negatives 0) and the True Negatives should be around 70 (and False Positives around 30), but this
heavily depends on the requriements of the system and the abilities of the used classifier.
A perfect classifier would have 900 True Positives, 0 False Positives, 100 True Negatives, 0 False Negatives
) Imagine you have to build an automatic insulin pump, a wearable device that keeps the blood glucose concentration BGC of a diabetes
patient in a healthy range of 4-5.5 mmol/L. Insulin is injected any time the BCG is higher than this range. A BCG lower than this range can be
fatal for the patient. The system stores in a log file all the insulin quantities that have been injected.
Identify 5 stakeholders for this system and formulate one functional and one non-functional requirement. [5p]
A stakeholder is anybody that has a direct or inderect influence on the product or is directly or indirectly affected by it.
Therefore, some of the stakeholders here are:
- the patient using the device
- the manufacturer
- regulators
- the companies producing the parts of the system
- lawmakers (i.e. the government)
Functional Requirement:
The system shall read the BGC of the patient
Non-Functional Requirements:
The system shall have an accuracy of 0.1 mmol/L when reading the BGC
The aim of Maneuvering Characteristics Augmentation System (MCAS) software is to maintain an airplane in horizontal position during
stalling, by pushing the nose down when the angle of attack is too high. For this software system,
Identify one accident; for that accident, identify one hazard; for that hazard, imagine one possible causal scenario and for this causal scenario
specify a mitigation measure to prevent the hazard from happening.
Finally describe a test scenario that can prevent this causal scenario from
happening. [10p]
Accident:
Hundrets of passengers are killed
Hazard:
The plane crashes
Causal Scenario:
The sensor measuring the Angle-of-Attack is broken
Mitigation Measure:
Install multiple sensors for measuring the Angle-of-Attack and add error checking code (e.g. comparing the values of all
sensors and giving an alarm in the case of a significant difference)
A possible test scenario in this case would be a simulation of the system during a stall, both with and without having a
“broken” sensor
