hello world and welcome. today, I will talk about a really interesting topic. I'm using another prophet to find some anomalies in a time series data. that's very interesting in um things like predictive maintenance. um where animal detection is used to find as early as possible some anomalies. yeah my name is Harris and let's get started. if you like the content please hit the subscribe button or leave a like comment there. first of all, I'm importing order required libraries like normal profit pandas numpy plotly and so on. then I'm loading the data which means I'm loading um one data which was collected from some sensors. I have um data frame um with some normal data and then I'm I use that to also put Four Peaks some kind of anomalies and the signal. so I'm using two um signals which I will show you here. um I have an about um blue original signal and the red is the same signal but with Four Peaks which I was inserted in the signal to make to simulate the animal is actually. then you can also plot the signal.

and with the power spectral density and a frequency spectrum. Which means here you have the time series data and the right, the frequency spectrum. Then you can see Peaks around about one or two Hertz, 10, 15, and 28 Hertz. The same I I was doing with the signal with the anomalies. Then you can also see that the frequency spectrum looks a little bit different, which you can see here when you compare them. You can see the red one is the signal with the anomalies, and you can see it's a little bit different like the original signal. Then I'm doing some either to plot everything, the signal with and this diagram. You can see it's going from 6 to 11 around about in the box plot. You can see it, um, there are small outliers here, a little bit going on. And then you can plot everything again as I have the signal here with, um, rolling mean, um, and then you have the standard deviation below where you can see if the standard deviation goes up, which means there's a

movement going on keep it in your mind. Then I'm training the Nora profit here with um um. I put only one forecast, which means I'm you using only one prediction or forecasting here. LX means I'm going only seven days back because in it's in a Time series data, it's based on days. Number of thin layers I put the two and number of hidden layers in the hard net is on four. So I'm using rnet in this case as well and the number of epochs I put only for this case for a small example to 10, but you can of course increase it to make your model a little bit better accurate as well. Then I'm splitting um the training and test data 20 to the test. Then I'm um trying everything. Then you can see the mean absolute error and mean at the root mean Square Arrow as well, which looks actually good. Then you can save and load the model here with pickle, saving and loading everything. They aren't doing the first prediction and also with training and test data.

then I'm plotting the metrics above. Which you can see here, the it's going down. The loss is going down because we have time series data. So we are talking only about losses. I'm calculating the anomalies, which the Delta from y hat one and why they can calculate it's or about 1, which means it's actually really good. Then I'm telling the model or the anomaly detection, if the Delta from the original signal and the predicted signal is higher than four, it's like a threshold. Then you are talking about anomalies. You can plot everything here. You can see the original data and the forecast, and you can really see here. Um, you can really see here some anomalies. And um, if you are plotting also the original data, then um, you have the training and the test prediction here. In the red, you have the training data frame and the test strata frame. It looks pretty good. And the red dots are the anomalies, which will find here. Um, you can also plot everything with the plot function from the nor profit. Here is the origin data, the black one.

and the blue one is um y hats of the prediction. You can also plot the parameter, select the trend within the years, the weeks, and so on. And yeah, I would say the prediction looks very good after 10-hour box. And yeah, thanks for watching um and see you soon.

Neural Prophet Anomaly Detection Predictive Maintenance | Python

Neural Prophet Anomaly Detection Predictive Maintenance | Python

Scripts