[kjd02002]

Dear TPRML,

We are investigating this issue and will provide an update ASAP.

[kjd02002]

Scores will not be provided or posted until after the completion of the competition.

[kjd02002]

Dear TeamBlue,

You may only submit results once for each stage. You will only receive data for the next stage after confirming the prediction you’ve submitted for the previous stage. This is to ensure teams do not use future data to inform their predictions.

[kjd02002]

Dear Arundites,

Thank you for highlighting this discrepancy. This is an error in the scoring spreadsheet. The competition will be scored with m = 10.

[kjd02002]

Dear ChoochooTrain,

With respect to your above questions:
1) Details regarding the final validation signal data will only provided during the validation period, starting July 14, 2019.
2) Crack length = 0 indicates that a crack has not yet initiated.
3) This information will not be provided.

[kjd02002]

Dear Apostov,

To win a data challenge prize, at least one member of the team must be registered and attend the PHM 2019 Conference.

Please note the following dates:
July 21, 2019 Competition Closed
July 28, 2019 Preliminary Winners Announced
August 11, 2019 Winners Announced

We will contact teams with the highest scores after July 21, 2019 to enable time for conference registration and travel planning before final winners are announced on August 11, 2019.

[kjd02002]

Dear binarysurface,

Thank you for highlighting this issue. I was able to replicate the problem while using Chrome browser. It was previously possible to download the data via Chrome so this is a relatively new issue. We will investigate and work to resolve this issue.

While troubleshooting, I was able to successfully download the data using Internet Explorer. As a temporary workaround, I recommend trying to download the data using a different browser.

Thank you.

[kjd02002]

Dear binarysurface,

The link you’ve posted above is not the correct address for the data sets. Please ensure that you have logged in, then go to https://www.phmdata.org/2019datachallenge/ and select “Downloads” from the subpage menu and you should see the files that can be downloaded.

[kjd02002]

Dear MaD_Lab,

Thank you for your question as they have identified 2 errors in the training data files. The data challenge coordinators recommend the following changes to the training data set:

Specimen T3: delete ultrasonic signal at cycle 55391
Specimen T4: change 7054 to 67054 in the Description file.
7054—>67054 2.74

[kjd02002]

Dear Duc,

There are multiple sensor pairs for each test sample. The sensor pair number is a unique identifier for each sensor pair. The System Description identifies the unique sensor pair associated with the crack initiation location for each sample.

[kjd02002]

Dear 20Years,

The information provided in the System Description is the level of detail available for this data challenge. More details maybe made available after the challenge is complete.

[kjd02002]

Hi ChoochooTrain,

Thank you for the example. I believe the discrepancy you’ve identified is that the penalty score is based on normalized crack length (% of final crack length) rather than absolute crack length (millimeters). For your example, this will result as follows: T=2+10*(7.46/7.46)=12, A=exp(|7/7.46-7.46/7.46|/0.2)-1=0.3611, M=1 –> S=T*A*M=4.333. This is not clearly identified in the penalty score description so we will update accordingly. Please let me know if this resolves your question.

Please note: normalized crack length is used in the penalty score calculation to ensure predictions for each sample has similar weighting regardless of final crack length. Teams will submit predicted crack length in millimeters (mm) and the penalty score will automatically be calculated based on the normalized crack length (% of final crack length).

Thank you.

• This reply was modified 4 years, 1 month ago by samuel_admin.

[kjd02002]

Dear TeamBlue,

You might be referring to a truncated waveform. If so, then this is due to the limit setting on the data collection instrument. This happens because the gain for actuation signal was adjusted in order to get the best waveform for the received signal (ch2) because that carries the information along the wave propagation path. For each specimen, the gain was kept constant through the whole testing process.

[kjd02002]

Dear ChoochooTrain,

Please provide an example of the discrepancy and we will investigate.

Thank you.

[kjd02002]

Hi Angler,

The answers to your questions are as follows:

(1) Data type: We understand the data type is ultrasonic. Is it correct?
– Yes, that is correct.

(2) Crack type: The specimens look like beams. We are asked to predict the fatigue crack length of the beams. Do you have any specifications of “crack length”, say the crack direction, crack depth and crack curve shape?
– The specimen is made of two aluminum plates mounted with 3 rows of countersunk rivets. Overall, the crack is close to a straight path, propagating from the edge of one countersunk hole to the adjacent one. Once the crack is seen on the surface, it is believed to be through the whole plate thickness.

(3) Data acquisition system: In our understanding, a lot of ultrasonic sensors are equally distributed on the specimen to pick up the ultrasonic signals. These signals are inputted to the receiver to be collected and further to a PC to be stored. Is it correct?
– Yes, that is correct.

(4) Test condition: When testing the ultrasonic signals at specific cycles, is the load added or not? We want to understand whether the load is applied all the time or not applied when conducting tests.
– When collecting the ultrasonic signals, the testing machine was paused near minimum loading to avoid introducing additional noise to the ultrasonic signal. The fatigue testing was resumed right after the ultrasonic data collection process.

[Author]

Posts