CSCI - 4146编程讲解、Data Science程序辅导、讲解c/c++语言编程 解析Haskell程序|辅导Database
- 首页 >> Python编程 CSCI - 4146 - The Process of Data Science - Fall 2020
Assignment 2
The submission must be done through Brightspace.
Due date and time as shown on Brightspace under Assignments.
● To prepare your assignment solution use the assignment template notebook available
on Brightspace.
● The detailed requirements for your writing and code can be found in the evaluation rubric
document on Brightspace.
● Questions will be marked individually with a letter grade. Their weights are shown in
parentheses after the question.
● Assignments can be done by a pair of students, or individually. If the submission is by a
pair of students, only one of the students should submit the assignment on Brightspace.
● We will use plagiarism tools to detect any type of cheating and copying (your code and
PDF).
● Your submission is a single Jupyter notebook and a PDF (With the compiled results
generated by your Jupyter notebook). File names should be:
○ A2--.ipynb
○ A2--.pdf
● Forgetting to submit both files results in 0 markings for both students.
Predictive maintenance (PdM) is gaining traction in the industry. In PdM, components are
replaced as they approach failure, not at prescribed intervals (Preventative Maintenance). For
PdM, equipment is monitored by sensors, and machine learning models are used to predict the
remaining useful life (RUL) (Fig 1.) of the equipment based on data streams generated by the
sensors. The data is typically a time series of sensor measurements collected until failure.
Figure 1: Illustration of an RUL.[1]
As shown in (Fig 2), a machinery health prognostic program is generally composed of four
technical processes, i.e., data acquisition, health indicator (HI) construction, health stage (HS)
division and RUL prediction. At first, measured data, such as vibration signals, are acquired
from sensors to monitor the health condition of machinery. Then, from the measured data, HIs
are constructed using signal processing techniques, artificial intelligent (AI) techniques, etc., to
represent the health condition of machinery. After that, according to the varying degradation
trends of HIs, the whole lifetime of machinery is divided into two or more different HSs. Finally,
in the HS which presents an obvious degradation trend, the RUL is predicted with the analysis
of the degradation trends and a pre-specified failure threshold (FT).[2]
Figure 2: Four technical processes in a machinery health prognostic program.[2]
In this assignment, you will need to predict an RLU of bearings. For the specific data set from
bearings (#4 of the datasets in the NASA Prognostics Center repository,
https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/), the data consists of
vibration measurements collected by accelerometers in the experimental set up (bearing test
rig) described in Fig. 16 of this publication and reproduced below. Each accelerometer provides
a single scalar measurement per sample. The sampling rate is 20KHz (20,000 samples per
second). The three-time series data sets are described in detail in this document. Each data set
consists of individual files containing 1-second worth of vibration signal measurements recorded
at specific intervals. The file name indicates when the data was collected. Each row in the data
file is a data point. Each row contains several measurements (channels), one from each
accelerometer in the experimental setup.
Example of a bearing (there are several other types).
1. Data understanding and feature engineering (0.1)
a. We will extract features from each channel of each of the data files of Test set 2.
The features will be statistical time-domain features typically used in bearing
monitoring. The six features to extract are RMS, Variance, Skewness, Kurtosis,
Shape factor and Crest factor (Table 1) [3]. Use σ=1 in the formulas for variance,
skewness and kurtosis. Your dataset should consist of 7 features: vibrational
signal plus the six time-domain features.
b. Build the data quality report
c. Identify data quality issues and build the data quality plan
d. Analyze your data. Plot the six features as functions of time for each of the
channels. Compute and plot the histograms of the vibration signals for each data
file. Describe your observations. How similar are the plots of the different
channels? Is there any evidence in the plots for which features are the most
useful for the RUL prediction task? Is the normalization of the data useful?
e. Preprocess your data according to the data quality plan
2. Build a baseline model to predict RLU (0.35). In Test set 2, there are four channels, with
channel 1 corresponding to the bearing that failed (bearing 1)
a. Explain what is the task you’re solving (e.g., supervised x unsupervised,
classification x regression x clustering or similarity matching x etc)
b. Use a feature selection method to select the features to build a model. Consider
different feature choices: features from channel 1 only, features from all four
channels, and different subsets of the six features.
c. Select the evaluation metric. Justify your choice.
d. Perform hyperparameter tuning if applicable.
e. Train and evaluate your model on test data from Test set 1
f. How do you make sure not to overfit?
g. Plot learning curve
h. Analyze the results
3. Build a NN model to predict RLU (0.35). Repeat question #2 above but now use a neural
network model to predict RLU. You can use a simple feedforward neural network or 1D CNN
from tutorial 6. Compare the model to your baseline model with a statistical significance test.
Use a box-plot to visualize your comparison.
4. Concept drift detection (0.2). Use concept drift methods and find out if there is any drift in
the data that can be detected. If so, what type of drift is that? Suggest specific actions to adapt
your model to the new concept.
References:
[1] D. A. Tobon-Mejia, K. Medjaher, N. Zerhouni and G. Tripot, "A Data-Driven Failure
Prognostics Method Based on Mixture of Gaussians Hidden Markov Models," in IEEE
Transactions on Reliability, vol. 61, no. 2, pp. 491-503, June 2012
[2] Machinery health prognostics: A systematic review from data acquisition to RUL prediction.
2018. Yaguo Lei ⇑ , Naipeng Li, Liang Guo, Ningbo Li, Tao Yan, Jing Lin
[3] Caesarendra, Wahyu, and Tegoeh Tjahjowidodo. "A review of feature extraction methods in
vibration-based condition monitoring and its application for degradation trend estimation of
low-speed slew bearing." Machines 5.4 (2017): 21.
Assignment 2
The submission must be done through Brightspace.
Due date and time as shown on Brightspace under Assignments.
● To prepare your assignment solution use the assignment template notebook available
on Brightspace.
● The detailed requirements for your writing and code can be found in the evaluation rubric
document on Brightspace.
● Questions will be marked individually with a letter grade. Their weights are shown in
parentheses after the question.
● Assignments can be done by a pair of students, or individually. If the submission is by a
pair of students, only one of the students should submit the assignment on Brightspace.
● We will use plagiarism tools to detect any type of cheating and copying (your code and
PDF).
● Your submission is a single Jupyter notebook and a PDF (With the compiled results
generated by your Jupyter notebook). File names should be:
○ A2-
○ A2-
● Forgetting to submit both files results in 0 markings for both students.
Predictive maintenance (PdM) is gaining traction in the industry. In PdM, components are
replaced as they approach failure, not at prescribed intervals (Preventative Maintenance). For
PdM, equipment is monitored by sensors, and machine learning models are used to predict the
remaining useful life (RUL) (Fig 1.) of the equipment based on data streams generated by the
sensors. The data is typically a time series of sensor measurements collected until failure.
Figure 1: Illustration of an RUL.[1]
As shown in (Fig 2), a machinery health prognostic program is generally composed of four
technical processes, i.e., data acquisition, health indicator (HI) construction, health stage (HS)
division and RUL prediction. At first, measured data, such as vibration signals, are acquired
from sensors to monitor the health condition of machinery. Then, from the measured data, HIs
are constructed using signal processing techniques, artificial intelligent (AI) techniques, etc., to
represent the health condition of machinery. After that, according to the varying degradation
trends of HIs, the whole lifetime of machinery is divided into two or more different HSs. Finally,
in the HS which presents an obvious degradation trend, the RUL is predicted with the analysis
of the degradation trends and a pre-specified failure threshold (FT).[2]
Figure 2: Four technical processes in a machinery health prognostic program.[2]
In this assignment, you will need to predict an RLU of bearings. For the specific data set from
bearings (#4 of the datasets in the NASA Prognostics Center repository,
https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/), the data consists of
vibration measurements collected by accelerometers in the experimental set up (bearing test
rig) described in Fig. 16 of this publication and reproduced below. Each accelerometer provides
a single scalar measurement per sample. The sampling rate is 20KHz (20,000 samples per
second). The three-time series data sets are described in detail in this document. Each data set
consists of individual files containing 1-second worth of vibration signal measurements recorded
at specific intervals. The file name indicates when the data was collected. Each row in the data
file is a data point. Each row contains several measurements (channels), one from each
accelerometer in the experimental setup.
Example of a bearing (there are several other types).
1. Data understanding and feature engineering (0.1)
a. We will extract features from each channel of each of the data files of Test set 2.
The features will be statistical time-domain features typically used in bearing
monitoring. The six features to extract are RMS, Variance, Skewness, Kurtosis,
Shape factor and Crest factor (Table 1) [3]. Use σ=1 in the formulas for variance,
skewness and kurtosis. Your dataset should consist of 7 features: vibrational
signal plus the six time-domain features.
b. Build the data quality report
c. Identify data quality issues and build the data quality plan
d. Analyze your data. Plot the six features as functions of time for each of the
channels. Compute and plot the histograms of the vibration signals for each data
file. Describe your observations. How similar are the plots of the different
channels? Is there any evidence in the plots for which features are the most
useful for the RUL prediction task? Is the normalization of the data useful?
e. Preprocess your data according to the data quality plan
2. Build a baseline model to predict RLU (0.35). In Test set 2, there are four channels, with
channel 1 corresponding to the bearing that failed (bearing 1)
a. Explain what is the task you’re solving (e.g., supervised x unsupervised,
classification x regression x clustering or similarity matching x etc)
b. Use a feature selection method to select the features to build a model. Consider
different feature choices: features from channel 1 only, features from all four
channels, and different subsets of the six features.
c. Select the evaluation metric. Justify your choice.
d. Perform hyperparameter tuning if applicable.
e. Train and evaluate your model on test data from Test set 1
f. How do you make sure not to overfit?
g. Plot learning curve
h. Analyze the results
3. Build a NN model to predict RLU (0.35). Repeat question #2 above but now use a neural
network model to predict RLU. You can use a simple feedforward neural network or 1D CNN
from tutorial 6. Compare the model to your baseline model with a statistical significance test.
Use a box-plot to visualize your comparison.
4. Concept drift detection (0.2). Use concept drift methods and find out if there is any drift in
the data that can be detected. If so, what type of drift is that? Suggest specific actions to adapt
your model to the new concept.
References:
[1] D. A. Tobon-Mejia, K. Medjaher, N. Zerhouni and G. Tripot, "A Data-Driven Failure
Prognostics Method Based on Mixture of Gaussians Hidden Markov Models," in IEEE
Transactions on Reliability, vol. 61, no. 2, pp. 491-503, June 2012
[2] Machinery health prognostics: A systematic review from data acquisition to RUL prediction.
2018. Yaguo Lei ⇑ , Naipeng Li, Liang Guo, Ningbo Li, Tao Yan, Jing Lin
[3] Caesarendra, Wahyu, and Tegoeh Tjahjowidodo. "A review of feature extraction methods in
vibration-based condition monitoring and its application for degradation trend estimation of
low-speed slew bearing." Machines 5.4 (2017): 21.