Are VAE used for missing data imputation in multivariate time series? If not, what is used? Multivariate time series are, to the best of my understanding, one of the few cases where Deep Learning still hasn't had its AlexNet moment. I'm especially interested to the case where most of the time series are continuous variables, with a few being categorical variables. To fix the ideas, think sensors of a big industrial machine - some sensors will record real-valued time series (pressure, temperature, speed, etc.) and some other categorical time series ("running/non running", "alarm 1/alarm 2/ no alarm", valve open/ valve closed", etc.). If the introduction of categorical variables makes the problem too hard, no issues - we can consider only continuous variables.
Some of these time series have a lot of missing data (the norm, rather than the exception, in industry) and I'd like to perform missing value imputation. In theory, a generative model seems the perfect fit for such a problem, thus I thought of VAEs. In practice, all the applications of VAE to missing data imputation that I know of, are related to images. Never seen them applied to missing data imputation for multivariate times series.
Can they be used? If so, which is the major modification I need to make to the architecture, in order to get a performing implementation for my use case? If no, which other Deep Learning or AI model are suitable for missing data imputation with multivariate time series?
 A: 
Can they be used?

Yes. Autoencoders (AE) are dimensionality reduction techniques. One could formulate a mapping from missing data series to full series, in 1-D, $\mathbb{R}^{m} \rightarrow \mathbb{R}^{n}$, where $m<n$, $m-n$ missing time-points. This is the conceptual idea. However, training vanilla AE may not be possible without introducing any prior knowledge from low-dimensional (missing data) set, that's why people introduce "pairing" AE with other techniques, such as Gaussian Process for the lower dimensional series $x_{t}$, to full latent series $z_{t}$. A sketch of the formulation in this case will look like  $$ p_{\theta}(z_{t}|x_{t}) = \mathcal{N}(g_{\theta}(x_{t}), \sigma^{2}\bf{I}))$$
where the inference is to find $\theta$ parameters. AE replaces function $g$. See GP-VAE: Deep Probabilistic Time Series Imputation.

If so, which is the major modification I need to make to the
architecture, in order to get a performing implementation for my use case?

The core issue is not only selecting appropriate architecture but formulation of the problem.

which other Deep Learning or AI model are suitable for missing data imputation with multivariate time series?

GANs are also utilised in the literature.
A: Wouldn't you consider language to be a type of time series? How about OpenAI Five's representation of the Dota game state as a time series with 20,000 continuous and discrete variables?
The tool of choice for such sequence modeling is LSTMs, Transformers, and other autoregressive models. And you can always tack on a latent prior to any of these models, (see "recurrent VAE"), but it's not necessary because any sequence distribution can be modeled as $p(x_0, x_1, \ldots) = p(x_0) \prod_j p(x_j|x_{i<j})$. 
The difficulty with imputation using these models is that you'd presumably be trying to do something like $\max_{X_i \ldots X_j} p(x_0 = X_0, \ldots x_{i-1}=X_{i-1}, x_i=X_i, \ldots, x_j=X_j, \ldots)$, and due to the autoregressive nature of the model, there's actually no easy way to carry out the maximization. If you just want to sample $X_i \ldots X_j$ from the modeled distribution with the other values fixed, there's no easy way to do that either. Maybe it would not be too expensive to use some form of MCMC sampling.
