Видео

@@oyewumicollinsikuejawa9488 yeah so it is an industry and academia collaboration. You are not only a student but also an employee to the company. Usually, certain days of the week you work at the company and certain days you stay at your lab at the University. The phd topics are more industry orientated as they are trying to solve a real life issue, which hopefully will help the company get an competitive edge. So the chances of you getting hired in the same company or a similar company after your phd is pretty high. Also, in france, doing a PhD isn't considered as relevant work experience but if it is CIFRE position, it could be counted as work experience too.

@Mohankumardash Thank you for this good information. I would be interested in such if possible for Finance, Accounting or Business Management related CIFRE PhD positions. I have MBA with specialisation in Accountancy.

Yes we can use all the engine's data to train the model. And that would be a better approach compared to the one shown in this video. I recommend first converting them into sliding window format and save as X (input) y(output) format and then start the training. And yes the same can be used for lstm model, you just need to change the model here

@@Mohankumardash Thank you so much for answering my question. I have another question: I have 10 time series data sets from different IGBT chips that are measured in parallel and have different lengths, e.g. one has 50,000 cycles and the other 30,000 cycles, because it's about the failure cycle of the chips. In what form can I feed this data into my LSTM- Regression model?

@@neuplatz4282 I had a paper published on the same type of problem: journals.sagepub.com/doi/abs/10.1177/1748006X221119301 But in my case the component was a capacitor and I had 20 time serise for 20 capacitors. Each timeserise coresponding to the life of a capaciroe. You can refer to this paper for more details.

I recommend not combining any other type of images while doing anomaly detection. So if you wanna do for pils then, just train using the normal data of pil images and evaluate on anomaly pil data. Train a different model for a new use case

@@Mohankumardash Thanks! Your videos very helpful!

That is also possible, but how will you encode this. Possibility is you will have to add extra layer with just 2 neurons between the 300 unit layer and the output layer. This bottle neck often results in sub optimal training. However with this if you can get good enough accuracy, then by all means you can plot the output from the layer having 2 neurons, and no need to use tsne.

Thanks for the query. You would be right if we were doing single variate analysis, however in this case we want to use 50 sensor data as input to lstm model, that's why we are choosing all the columns except first 3

@@Mohankumardash Thanks. I mean you explained that we will consider the n+1 the label for a sliding window of n, however in the code its considering n th label. Also, the code clearly commented we are considering the label for last row of the window, which is in line with the code. Not a big thing that changes the results, but a small confusion.

I am glad you find it useful, I will keep bringing on this sort of content. Nice username by the way

Hi, I saw your query. And if you have multiple classes to detect, it becomes a classification problem. And that video was for anomaly detection, it is only useful if the you want to detect whether the new image belongs to the known distribution or not. You can also use the anomaly detection model as classifier too. But then you should have N number of anomaly detection model each representing images belonging to specific class, here N is total number of classes. Or the better approach will be to use a classifier instead, like shown in this video. I hope that helps.

@@Mohankumardash Thanks for the response and I actually tried multiple anamoly detection models but when I tried to predict on test image and run it through multiple anamoly detection models atleast 2 to 3 anamoly detection models are matching to the known distribution with low anamoly score and the diff between the score is <0.05 do you have any suggestions which classifier or algorithm is best suited for multi class classification?

@@Mohankumardash I saw the video and it is using mobile net pre-trained model, for mvtec dataset can we use the similar idea of using hooks(mentioned in anamoly detection video) and aggregate the layers(1536 dimensions) and avoid last layer(2048) and then train the model for better performance. btw Great explaination on this video

@@kidsstories-cw2yc Yes, that is definitely possible. Just like for the mobile scratch case, I had 3 classes, you can first define the classes you have for your dataset and divided them into subfolders. Then you can easily fine-tune the pre-trained model (mobile net or anything else) for your dataset. There is another way, too. You can use the pre-trained models as feature extractors and get the 1536-dimensional vector embedding for each image. Then, train a basic classifier like SVM, decision tree, or Random forest for the classification task. However, in my opinion, this approach tends to give subpar results compared to fine-tuning the whole model.

@@Mohankumardash ok Thanks and do you have the tutorial video on above suggested approach? , any video which approximates the above suggested approach is also fine to me

Hi, I am glad you liked the video. I would say there are better methods to detect misalignment, such as Patchcore or efficientAD. I created a video on Patchcore too: ruclips.net/video/lOFv59Hvr50/видео.html

Thanks!!!

I am glad you find it valuable

@@Tourdeglobe hello, Thank you for reaching out. But I don't think I have any videos specific to nuclear power plant fault diagnosis. The closest video you can find is on fault diagnosis of industrial process: ruclips.net/p/PLoSULBSCtoffIldbr898SDp5gIqo8XL-t&si=nbh04Qq05yOP4fmT

@@Tourdeglobe hello, Thank you for reaching out. But I don't think I have any videos specific to nuclear power plant fault diagnosis. The closest video you can find is on fault diagnosis of industrial process: ruclips.net/p/PLoSULBSCtoffIldbr898SDp5gIqo8XL-t&si=nbh04Qq05yOP4fmT

@@Mohankumardash Thank you sir

Sorry for the confusion. It is the same file that was created in the previous video, but it has been renamed. Important thing is it should have the RUL column. It is obtained by subtracting current cycle from EOL cycle

@@harshitakhatri9002 In France there are two types of job contracts, CDI and CDD. CDI is for permanent position and CDD is fixed duration position. Usually CDD are easier to find, especially in the form of Postdocs in and around France. But the issues is you will need to renew it every year. If the lab doesn't want to renew it you will have to look for another opportunity. So your are always on the edge of uncertainty. For CDI, there are public position like becoming a Professor or Researcher in university or labs. But here the competition is very high, usually you will be competing with people with 4-5 years of postdocs and 20-30 journal publications. Finally comes an industry job, where you need a good french fluency and a relevant company who values your research. To sum it up, there will be struggle but maybe not as much as you would have to do in india after a PhD

@@lab-test2601 Hi, to calculate the residual signal first you have to obtain an redundancy relation for your system dynamics. This is the fancy way of saying, you will have to find the governing equation such as Kirchhoff's law, newton's law, Bernauli's law etc. If you have a simple L, R circuit with a voltage source of V. The ARR: V - L di/dt - R i = 0 In this equation you know the parameter values such as L, R and the input source V. Then bu using one current sensor you can obtain the 'i' and plug in this equation with time to get the residual signal. Ideally it should be zero if the system is in healthy state. But even in healthy state you get little bit deviation because of the temperature, humidity etc that influence the parameter values. If in your system, the R is faulty (if it's value increases or decreases than nominal) then the right hand side of ARR won't be zero anymore, that's how you detect the fault

@@Mohankumardash Thank you! so much. di/dt is zero or ?

@@lab-test2601 welcome 😁. No not always you have to calculate it everytime by i_(t+1) - i_(t)/ ∆t

Absolutely I will !

Soon soon

Thank you very much for your support 😊

@@AmpedUpWithKrishna , Thanks for your question. I encourage you to look into the first video of the series to know the details about the dataset and how they are extracted: ruclips.net/video/iCTU-IZ6rPQ/видео.htmlsi=sW3XTc79aECBq4LM.

@@Mohankumardash ok Thank you!

I am happy to know it helps you

Hi, thanks for the questions. So for if one feature in your model is constant throughout, then it'll not have any discriminative power. This means the standard deviation is zero (no variance) for this feature. We want those features, who have a high variance to train a model. This gives an indication that, this feature changes when the fault condition changes (it's not always true). But it is a good starting point to ignore the features having very low or zero standard deviation

@@Mohankumardash alright understood thank for clarifications😃

@@lucysii6085 I am glad 😊

Hi Suresh, I am glad you liked it. I already uploaded a few videos on patchcore. You can check them out, and one more video will be released today on patchcore. I am very glad you found them interesting, I will keep posting new videos about the state of the art algorithms. I request you to share my videos on LinkedIn or with your peers if you think, it will help them.

Hi Gajendra, thanks for your feedback. I would like to ask you, what more I can add to these videos to make it an end to end project. As I am an AI researcher, I mostly dela with model training and trying out new methods. According to you, what would make this an end to end project?

@@MohankumardashI appreciate the efforts on channel, you have a lot of knowledge about AI and ML. i was talking in general like a playlist where you can put videos in which one can see diff end to end projects. although your content is nice

Thank you, they will be uploaded soon.

Thank you very much

Definitely. soon I will upload new videos on my thesis.

I don't think so. Because the pertained resnet model is trained on image data, it won't perform good for bearings

@@Mohankumardash Thank you. I watched the CNN 3D model video. I use it to generate 300*300 image dataset with PMSM engine and process it with image CNN model. I find it effective. I see this video converts grayscale images to color images. So I don't know if I can or not

@@shenjgaming2988 I can't say for sure but you should try it out and see if it works or not. All the best

@@Mohankumardash Thank you sir. Your videos have helped me a lot

Thank you. I am currently working on such models, when they will be ready, I will publish them, definitely 😁

Glad you think so!

Thanks!

Thanks a lot