Hello Sdi,

SVM evaluation can be improved. But, before that the problem needs to be clearly defined. Feature selection based on your domain may increase the SVM performance also. In addition, parameter tuning is also a fair choice, indeed.

Thanks,

Sobhan

Thank you Alex for your replies.

Yes I normalized data with this command:

C=10;

svmstruct=svmtrain(TrainInputs,TrainTargets,...

'autoscale',true, 'boxconstraint',C,...

'kernel_function','rbf',...

'rbf_sigma',4 );

At first I used 80% of whole data for training (19202*6) and 20% of whole data (4800*6) for test, with 2 classes include normal and fault , for the second time my features reduced to one.

I used SVM toolbox , svmtrain for training and svmclassify for test.

Respectfully,

Dear Faeze Sdi ,

To improve the performance of the system I suggested you to do the first feature normalization then feature reduction then feature selection to select the optimal feature from whole feature vector and passing this feature to SVM for classification , but before to passing that check the configuration parameter of SVM.

Dear Mouad M.H. Ali ,

Thanks for your reply. Would you please explain more about what do you mean by this sentence: " check the configuration parameter of SVM "?

Do you mean C and rbf_sigma?

Respectfully,

Yes

Dear Faeze Sdi,

You could try to tune your SVM:

http://ieeexplore.ieee.org/document/7280664/

http://ieeexplore.ieee.org/document/7280644/

Best Regards,

Sylvio.

Dear Sylvio,

Thank you, I will read papers.

Respectfully,

The results from grid searching show you whether your results can still be improved or not. For example, when I tune the parameters of an SVM classifier, I first start with the kernel choice. When I see that a quadratic kernel performs better than its linear counterpart, I check whether a cubic kernel achieves an even better performance. If this is not the case, one can reasonably infer that a quadratic kernel is the best choice for the given dataset. Subsequently, I use the same approach for the trade-off parameter C. From my personal experience, this approach does a good job in tuning parameters.

However, when optimising your algorithm for the available data, you should be aware of two things:

The problem of overfitting and computational complexity show that parameter tuning ensues trade-offs. When you want to know whether further optimisation of the parameters is required, it may help to ask yourself the following questions:

What is the purpose of the algorithm? Is it correctly classifying as many instances as possible or having a high recall rate for the first class? Does further tuning help the algorithm to do better in this task? Does the current choice of parameters allow for training the algorithm in a reasonable amount of time?

Dear Mazin Alshujeary,

Thanks for your explanations, it was useful. I searched a lot about

grid search for tuning parameters of SVM in MATLAB, but I didn't find anything .

Would you please explain a little about grid search ?

Respectfully,

sklearn provides you the library named GridSearchCV. You can perform pre-processing using sklearn to obtain the best results.

Thank you Hussain Ahmad.

Thanks Huseyin Guney for your explanations , actually these results are after normalization.

I'll try to use k-fold Cross Validation.

If you have less accuracy using linear SVM. Try to do it with quadratic SVM and cubic SVM. These are also available in SVM tool box.

Because of complex data linear SVM some time not able to efficiently classify them.

Some important things to do when using SVM:

- Scale/Normalize data;

- Perform a grid search for best parameters;

- If data is unbalanced do not use accuracy for measuring performance, choose something like f1-score ou area under ROC curves.

Faeze Sdi,

Grid search is nothing but trying every possible combination of parameter values to check the optimal result. The parameters can be explored quickly by changing them by a constant step and by using a log function at every local optima for fine tuning. Grid search implementation is not that hard with matlab. You may check libsvm which provides a tool for parameter optimization with grid search.

If the training takes less time, you can try optimizing the parameters exhaustively. If you still cannot get the required tp and tn then your feature set might be not discriminative enough for the dataset.

I suggest to check the eigenvalues once the PCA performed over the dataset to know whether the selected principal components are strong. Also try with other known handcrafted feature set.

Hi,

For tuning the parameters such as C and Gamma its better to use the LibSVM library.

Hello Faeze Sdi,

I can see TPR is around 70% . Still I am not wonder with this accuracy. It is true that you have potential opportunity to improve the classification accuracy. I can see your overall configuration of SVM is correct, however you can refer follow points

1. Apply grid search to tune parameter C and Sigma( for RBF Kernel). You can follow this article Chapter Multi-fault Diagnosis of Roller Bearings Using Support Vecto...

2. Did you try with other algorithms and what happens to them?

3. I can see that you are using only 6 features, this may not give you good result. Better try with more robust feature extraction process as bellow, Article Reliable Multiple Combined Fault Diagnosis of Bearings Using...

4. You have benchmark dataset that available in online. You can apply same algorithm and check what happens,

http://csegroups.case.edu/bearingdatacenter/pages/download-data-file

Hope for the best,

Regards,

M M Manjurul Islam

South Korea

Hi,

Thank you guys for your replies.

Unfortunately I can't change methods . I have to use SVM ,SVM+PCA.

I didn't try other algorithms with my data set. I picked 6 features because when I created fault I could see changes in 6 features more than others.

I normalized data before SVM and PCA+SVM , result became better but still they aren't best.

I will try to use k-fold Cross Validation or grid search.

Hello Sdi,

SVM evaluation can be improved. But, before that the problem needs to be clearly defined. Feature selection based on your domain may increase the SVM performance also. In addition, parameter tuning is also a fair choice, indeed.

Thanks,

Sobhan

Hello Sdi,

The problem may be your features, you should get new features or generate more features. I also recommend applying Pearson correlation to eliminate redundant features. You can read my paper to improve the idea. Any questions you can send me a message.

Multi-fault Diagnosis of Rotating Machinery by Using Feature Ranking Methods and SVM-based Classifiers (DOI: 10.1109 / SDPC.2017.29)