🤖 machine learning

A Scalable Multi-Task Model for Virtual Sensors

This paper introduces a scalable, multi-task machine learning model for virtual sensors that simultaneously predicts diverse signals with improved efficiency and accuracy by leveraging task synergies and automatically learning relevant inputs, thereby overcoming the computational and scalability limitations of existing isolated or foundation-based approaches.

Original authors: Leon Götz, Lars Frederik Peiss, Erik Sauer, Andreas Udo Sass, Thorsten Bagdonat, Stephan Günnemann, Leo Schwinn

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Leon Götz, Lars Frederik Peiss, Erik Sauer, Andreas Udo Sass, Thorsten Bagdonat, Stephan Günnemann, Leo Schwinn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the engineer in charge of a massive, high-tech factory. This factory has thousands of sensors measuring everything: temperature, pressure, speed, and voltage. But there's a problem: some of the most important sensors are incredibly expensive, break easily, or simply don't exist yet.

Traditionally, to solve this, engineers would build a separate, custom-made robot for every single missing sensor. If you needed to predict the temperature of a battery, you'd build Robot A. If you needed to predict the speed of a wheel, you'd build Robot B.

The Problem: This is like hiring a different chef for every single dish on a menu. It's expensive, slow, and the chefs don't talk to each other. If Robot A learns something about heat, Robot B doesn't know about it. Also, you need a human expert to tell every robot exactly which other sensors to listen to, which is hard to get right for thousands of robots.

The New Solution: The "Super-Brain" Chef

The authors of this paper propose a different approach: One Super-Brain Chef that can cook any dish on the menu, instantly.

They built a single, unified AI model that can predict hundreds of different "virtual sensors" (the missing measurements) all at once. Here is how they made it work, using simple analogies:

1. The "Empty Plate" Trick (Selective Prediction)

Usually, AI models are like students who can only repeat back what they just heard. If you ask them to predict a new signal they've never seen, they get confused.

The Innovation: The authors gave their model a special trick. When you ask it to predict a specific sensor (like "Battery Temperature"), it starts with an empty plate (a zero-vector). It then puts a special label on that plate saying, "This is for Battery Temperature."
How it works: The model looks at all the available data (the ingredients) and figures out which ones are relevant to fill that specific plate. It learns to ignore the noise and focus only on the signals that actually help predict that specific temperature. It's like a chef who, when asked for a soup, instantly ignores the spices meant for the steak and grabs the right vegetables.

2. Learning Who to Listen To (No Expert Needed)

In the old way, a human expert had to sit down and say, "Robot A, listen to Sensor X and Sensor Y."

The Innovation: This new model learns on its own who to listen to. It has a built-in "attention mechanism" that acts like a spotlight. During training, it figures out which input signals are important for which output.
The Result: It creates a unique "listening list" for every virtual sensor. For a battery temperature, it might listen to coolant flow. For engine torque, it listens to pedal pressure. It does this automatically, removing the need for a human expert to hand-pick the inputs.

3. The "Smart Filter" (Efficiency)

You might think, "If one model does everything, it must be huge and slow."

The Innovation: The model is incredibly efficient because of its "Smart Filter." When you ask it to predict just one sensor, it doesn't look at all the factory data. It uses its learned "listening list" to ignore 99% of the inputs and only process the few that matter.
The Analogy: Imagine a librarian who has to find a specific book. Instead of walking through the entire library, they know exactly which shelf and which row to go to. They skip the rest. This makes the model 415 times faster and use 951 times less memory than trying to run hundreds of separate models.

The Results: A Factory That Runs Itself

The team tested this "Super-Brain" on real-world data, including a massive dataset from an electric car (with over 18 billion data points) and standard traffic/electricity datasets.

Scalability: They showed that adding more virtual sensors (from 1 to 512) didn't make the model much bigger or slower. It's like adding more dishes to the menu without needing to hire more chefs; the one chef just gets better at multitasking.
Quality: The predictions were just as good, or even better, than the old method of using separate models for each sensor.
Explainability: Because the model learns its own "listening lists," we can look at it and say, "Ah, to predict the battery temp, it is listening to the coolant flow." This makes the AI transparent and trustworthy, which is crucial for safety-critical things like cars.

Summary

Instead of building a separate, expensive, and isolated robot for every missing sensor, this paper introduces one flexible, efficient AI that can predict hundreds of missing signals simultaneously. It learns which data to ignore and which to focus on, saving massive amounts of computing power while making the system easier to understand and scale. It turns a chaotic factory of thousands of individual robots into a single, highly organized, and super-fast team.

Technical Summary: A Scalable Multi-Task Model for Virtual Sensors

Problem Statement

Virtual sensors utilize machine learning to predict target signals from available measurements, offering a cost-effective and reliable alternative to expensive physical sensors in critical applications like autonomous driving and healthcare. However, existing approaches suffer from four primary limitations:

Lack of Scalability: Current methods rely on isolated, application-specific models trained and maintained separately for each virtual sensor, with input signals hand-selected by domain experts. This prevents scaling to large sensor networks.
Missed Synergies: Isolated models cannot leverage task synergies among different virtual sensor tasks.
Inconsistent Benchmarks: There is a lack of consistent evaluation standards, with few works utilizing standard benchmark datasets.
Foundation Model Limitations: While time series foundation models offer general-purpose solutions, they are computationally expensive and limited to forecasting input signals. They cannot predict new, unseen signals from other measurements, rendering them incompatible with virtual sensor tasks.

Methodology

The authors propose a unified, scalable multi-task model based on a causal decoder-only transformer architecture. The core innovation lies in three mechanisms designed to exploit task synergies while maintaining computational efficiency:

1. Selective Prediction of Virtual Sensors

To predict virtual sensors (signals not present in the input set), the model employs an autoregressive mechanism using prototype tokens.

Tokenization: Input time series are normalized, patched, and embedded into tokens.
Prototype Insertion: To forecast a specific virtual sensor $j$ , a zero-vector prototype token is inserted into the input sequence.
Variate Embedding: A learned variate embedding vector $v'_j$ is added to the prototype token. This guides the neural network to map relevant target information into the empty prototype, allowing the model to predict the specific virtual sensor based on available input signals.
Autoregression: The model predicts the next $p$ time steps. In subsequent steps, the model's own previous prediction serves as the new input for the virtual sensor, mimicking real-world deployment where physical inputs are available but virtual outputs must be predicted.

2. Learning Relevant Input Signals

Instead of relying on expert-selected inputs, the model learns the set of relevant input signals for each virtual sensor end-to-end.

Signal Relevance Vectors: The model introduces trainable signal relevance vectors $R' = \{r'_j\}$ for each virtual sensor. Each vector encodes the relevance of all available sensors (inputs and other virtual sensors).
Attention Bias: The relevance vectors are used to compute a cross-relevance matrix via an outer product, which is duplicated across time and applied as a static bias in the attention score computation of every transformer layer. This allows the model to learn global, time-independent signal importance.
Explainability: By analyzing these vectors, the model provides explainability regarding which input signals drive a specific virtual sensor prediction.

3. Efficient Sparsification and Selection

To address the computational constraints of large sensor networks, the model employs structured sparsity:

Thresholding: During late training stages and inference, signal relevance vectors are sparsified based on a threshold $r'_{thres}$ . Unimportant signals are masked with $-\infty$ , effectively removing them from the attention mechanism.
Dynamic Sensor Selection: The model can flexibly select which virtual sensors to predict in a single generation step. When predicting a subset of sensors, the model queries only the relevant input signal sets for those specific sensors, minimizing the number of tokens processed.
Training Strategy: To avoid identical gradients for all sensors (which would lead to identical input sets), the model trains on a random subset of virtual sensors ( $N_{train}$ ) per iteration. Additionally, Teacher Forcing is used during training to reduce memory and computational costs, with a hybrid approach (fine-tuning with Backpropagation Through Time) explored to mitigate error accumulation.

Key Contributions

Large-Scale Benchmark: The authors provide the first large-scale analysis of unified virtual sensor models on three standard time series forecasting benchmarks (Traffic, Electricity, Solar) and an application-specific automotive CAN bus dataset containing over 18 billion time series samples.
Unified Multi-Task Architecture: The proposed model simultaneously predicts diverse virtual sensors, exploiting task synergies and foundational time series knowledge. It scales gracefully to hundreds of virtual sensors with a nearly constant parameter count.
Learned Relevance and Explainability: The model eliminates the need for expert knowledge in input selection by learning relevant input signals end-to-end, thereby adding explainability to the virtual sensor design.
Efficiency Mechanisms: Through selective prediction, learned relevance masking, and sparsification, the architecture achieves significant reductions in computation time and memory usage.

Results

The evaluation demonstrates substantial efficiency gains while maintaining or improving predictive quality:

Performance vs. Isolated Models: Compared to isolated models trained for single sensors, the unified approach generates superior predictions at similar inference speeds.
Efficiency Gains: On the CAN bus dataset, the unified model reduces computation time by up to 415× and memory requirements by 951× compared to standard multi-task approaches that do not utilize sparsification or selective prediction.
Scalability: The model maintains stable predictive quality (low standard deviation in MSE) when scaling from a few to hundreds of virtual sensors, whereas isolated models would require a linear increase in parameters.
Explainability: Qualitative assessment on the CAN bus dataset confirms that the learned input signal sets align with domain expert knowledge (e.g., temperature sensors relying on other temperature and flow rate signals).
Transfer Learning: The model can efficiently introduce new virtual sensors into an already trained model using only 10% of the training data, achieving performance comparable to training from scratch.

Significance and Claims

The paper claims to present the first unified model for virtual sensors that successfully bridges the gap between application-specific virtual sensors and time series foundation models. Its significance lies in:

Practical Deployment: By drastically reducing computational and memory requirements, the model enables the deployment of virtual sensors on edge devices and in large-scale sensor networks where current foundation models are too resource-intensive.
Elimination of Expert Bottlenecks: By learning input relevance automatically, the approach removes the dependency on domain experts for feature selection, making virtual sensor design more scalable and robust.
Synergy Utilization: It demonstrates that virtual sensor tasks can benefit from shared knowledge, improving predictive quality and efficiency simultaneously.

The authors conclude that this multi-task model unlocks new applications for virtual sensors and contributes to more environmentally sustainable time series modeling by reducing the computational footprint of large-scale sensor networks. They acknowledge limitations regarding hyperparameter search due to the high cost of training large-scale models, suggesting potential for further optimization.