The tales of Llaima, the volcano
I receive emails from students asking for an opportunity to work with Artificial Intelligence(AI)/Machine Learning(ML). However, when I present some projects related to signal processing, I see the sunshine fading away from their eyes... Why?
I know these machine learning vibes (decision trees, SVMs, CNNs, Boosting, etc.) on i.i.d. datasets are exciting. I'm also surfing such waves! But, what happens to the general perception when the datasets have some temporal dependencies? I've attended conferences supported by the IEEE Signal Processing Society, where I'm also a member, and other researchers mentioned the same difficulties. A researcher told me once: "Never start a conversation with a student mentioning the signal processing area, but AI, ML, Deep Learning,...". That's why the title of this post mentions "Volcano", but honestly, I'd like to share some papers, codes, and datasets connecting Machine Learning and Signal Processing to show how impressive this area is!
Let's start with the dataset. From a great collaboration with Prof. Millaray Curilem, a researcher from the Universidad de La Frontera (Temuco - Chile), we organized four different seismic signals collected from the Llaima volcano into a dataset available here:
Llaima is an active volcano in Chile, specifically in the Araucanía region - the southern Andes. This volcano is part of a chain on an active tectonic boundary between the Nazca oceanic plate and the continental plate of South America.
The dataset assembles signals from different sources: Volcano-Tectonic (VT), Long Period (LP), Tremor (TR), and Tectonic (TC). VT refers to the brittle failure of rocks inside the volcanic building, which is the same type of event that happens along purely tectonic faults. It presents a frequency pattern with a broadband spectrum that may reach 10 Hz. LP corresponds to the resonating of magma and gases inside volcanic conduits toward the surface, whose spectral pattern is narrower than VT, mainly bounded in [0.5, 5] Hz. LP events are usually observed before volcanic eruptions. TR is a continuous and high-amplitude signal produced by several processes, such as long-lived resonance due to the extended flow of magma movement through cracks. Their broadband spectrum is usually in the range [0.5, 3.0] Hz, slowly attenuating at the event's end. At last, TC events are not related to volcanic activities, typical of the dynamics of geological faults. TCs may result from local, regional, or even distant activities regarding the epicenter location.
In our seminal study, we have used fundamental signal processing tools to model those signals. It was a very interesting work designed to discriminate volcano seismic signals using spectrogram cross-correlations (SPCC).
For those unfamiliar with spectrogram analyses, I shared a Python notebook to illustrate how Fourier can transform a Llaima signal from the time to the frequency domain, producing spectrogram and PSD (Power Spectral Density) plots - fundamental concepts for this first paper. Check it out:
In this first manuscript, we also used Machine Learning to take advantage of using SPCC as a distance measure to compare two signals. So, an Instance-based Learning method works as a natural choice to discriminate these volcano sources.
Next, we have decided to assess whether spectrograms produced by Short-Time Fourier Transform could be used to train Deep Neural Networks (DNNs). Yeap!, the results were pretty amazing:
I have also shared Colab notebooks for two DNNs trained on the Llaima signals. The goal of this notebook is not to extract the best possible performance (check our paper for this) but to illustrate didactically the usage of DNN in this context.
In the first example, we have trained a Multi-layer Perceptron (MLP) on spectrograms transformed from two different types of Llaima signals:
In the second one, we have trained a CNN (1D) directly on the Llaima signals!
(Again, the experiments run a few epochs and folds just to illustrate our experiments)
Next, we analyzed several DNN architectures and verified other signal transformations. The best results ever were obtained using Continuous Wavelet Transforms (CWT)!
领英推荐
Ok, the results are thrilling, but how do the models learn? Once we're using black-box models, this is a tough question. Nowadays, there are several concerns about interpretability and explainability issues. So, we decided to use a very interesting tool called Layer-wise Relevance Propagation (LRP) to travel through the DNN layers. In summary, LRP is a visualization method used to identify the importance of pixels through a backward pass in DNNs, in which neurons that contribute the most to the higher layer receive greater relevance.
In that work, we have used LRP to assess the influence of noise on the Llaima signals. The next figure shows how the noise level changes the DNN "perception" of the signal sources.
And guess what?! Once I can see the region that positively affects the DNN (red pixels) and the negative region (blue pixels), why not combine interpretability and filtering as a denoising tool?
Since my PhD, signal separation is a theme that moves me a lot, as you can see in the following papers (no volcano signal, sorry):
In Machine Learning, it is usually necessary to preprocess the data before training any model. A common preprocessing step is removing noise to avoid the problem mentioned in Fig. 4, in which the noise affected the DNN performance.
Of course, we could use the denoising tool that combines interpretability and filtering. But we challenged ourselves to skip this step by creating a DNN architecture that is somehow more robust to noise! In the following manuscript, we have combined CNN and LSTM to implement a cell that's not so affected by the noisy influence.
We have noticed that, using this new LSTM cell (NTSE-LSTM), the classification of some sources, such as LP, was not strongly affected by noise.
The source codes used in the experiments are also available for those interested in collaboration. I see several interesting points to improve these results.
In our most recent paper, we presented a methodology for developing a volcano-seismic event classification system using a multi-station deep learning approach to support monitoring the Nevados del Chillán Volcanic Complex, which has been active since 2017. Work designed in Chile that I was kindly invited to collaborate with.
That's?not?all, folks, fortunately! But, unfortunately, I cannot tell the spoiler of our next work that will be available soon. I hope to drop by here soon to tell another volcano tale... I mean, another signal processing research.
Special thanks to Prof. Millaray Curilem and my former students Jo?o Paulo Canário and Otávio Ribeiro.