Teng Zhang

Huazhong University of Science and Technology | hust · State Engineering Research Center of Digital Manufacturing and Equipment

We know that robot pose errors are very common and that robot trajectories are pre-constructed and generated from CAD models in machining scenarios. However, for the trajectory points, there are inherent errors, and we need to compensate for the position and attitude of these trajectory points.

The existing off-line compensation is very common, but it lacks real-time, and the compensation objects are all the results obtained from pre-experiments. In fact such compensation, in a new experiment, the whole process is different from that of the pre-experiment because of the compensation done, so the final compensation also all stay at the level of being able to improve the performance, and theoretically such compensation is also all incomplete.

My robot is an ABB, and it would be great if you could offer some advice on the robot control system,transmission of data, branching of the perception model, etc.

Thanks to all the researchers who discussed and replied.

Does clustering contain some potential knowledge at the data level? In terms of sense, clustering seems to find a measure of distance similarity between data. For classification problems, a class of data usually has similar characteristics. So what does it mean to find the edge cluster of a set of data using the elbow method for regression problems?

I really want to get:

Thank you all.

As I am currently doing regression related work, I hope to have some open data and help me to test my model, so I need you to provide me with some datasets that can be used. The requirement of datasets is N *feature*time series, where N represents the number of samples, feature represents the number of original feature channels.Time series refers to the number of channels sampled. At present, I have NASA's WEAR dataset, but I think it is not enough, so I need your help.

When evaluating regression model prediction effect, why usually usually is to use single evaluation index, such as R2 or MAE, RMSE, why not use a few of them together, for example using RMSE/R2 as evaluation indexes, namely when the R2 is 1 RMSE, do so whether there is a potential problem and hope to get your help

I would like to investigate the existing measuring-processing integration software, sincerely hope that you can share their own information, the more detailed the better

When I use a machine learning algorithm to make regression prediction, why do I find that there is a constant between the predicted value and the actual value? Could you tell me what may cause this phenomenon?Is it a model problem or a data problem?

In thin-walled milling experiments, I collected the whole process of cutting force signal, due to my final research is to use the type value points of the information on the thin wall piece, so I have to type cutting force signal and the value information match point, so I should be on the cutting force signal is split, excuse me I how to split the cutting force signal, in order to ensure better corresponding relation.In addition, since I subsequently adopted a deep learning model, I need to normalize the data in advance. I would like to ask whether I should first normalize the overall cutting force signal and then split it or normalize each segment after splitting it.

For the 100 measuring points in the figure, add the cutting force values of 1 million moments in the processing process that I have collected, and how can I match the values of measuring points (such as deformation) with the cutting force at this point?

As we know, in the research of machine learning or deep learning, training set and test set should be divided. The work of the training set is to train the model, while the function of the test set is to test the model on the trained model.However, I have seen some articles that test a model using the training set itself in order to demonstrate the effect of a model. Is there a problem with such logic?