Time series classification is a fundamental task in diverse fields such as medical signal analysis, industrial monitoring, and sensor-based activity recognition. In these contexts, class information manifests through localized shapes, specific frequencies, temporal shifts, or complex cross-channel interactions. Traditional approaches based on deep learning require large volumes of labeled data and costly training, while manual feature engineering demands expert domain knowledge. An efficient and robust alternative is random convolutional transforms, such as Rocket, MultiRocket, and Hydra, which convert variable-length sequences into fixed-dimensional tabular representations with low computational cost. However, until now these features were typically combined with simple linear classifiers, limiting their ability to capture complex patterns.
In this article we explore a new approach that integrates these convolutional representations with pretrained tabular foundation models. This combination, which we call MASHT (MultiRocket And Hydra with Tabular foundation model), enables direct classification without task-specific training. The process is remarkably simple: first extract features using MultiRocket and Hydra, then feed them into a tabular foundation model that learns in context, i.e., infers the relationship between features and labels from the provided examples. This completely bypasses the training phase of a custom classifier, accelerating development and facilitating deployment in environments where data changes frequently.
Experimental results with MASHT show that it matches or surpasses the most advanced baselines in univariate classification, achieving a lower average rank than HIVE-COTE 2.0, one of the most powerful classifier ensembles. In multivariate tasks, performance remains highly competitive against the strongest reference methods. This opens the door to real-time applications and systems that require continuous adaptation without human intervention. For example, in an industrial plant, sensors generate time series of temperature, vibration, and pressure; with MASHT, an anomaly can be detected instantly without retraining models every time a new sensor type is added or operating conditions change.
From a business perspective, adopting techniques like MASHT represents an opportunity to optimize processes and reduce costs. At Q2BSTUDIO, a company specialized in software development and technology, we understand that the key is to integrate these capabilities into complete and customized solutions. For instance, a time series monitoring system can benefit from custom software applications that incorporate tabular foundation models for real-time classification. Additionally, cloud infrastructure is essential to handle the data volume and required latency; therefore, we offer cloud AWS and Azure services that guarantee scalability and security. Cybersecurity also plays a critical role, as industrial or medical sensor data is sensitive; our solutions include cybersecurity and pentesting to protect information integrity.
Artificial intelligence is the engine driving these innovations. At Q2BSTUDIO we develop AI applications, including AI agents that can interact with time series classifiers to make autonomous decisions. For example, an AI agent could receive the MASHT prediction and trigger an alarm or adjust process parameters without human intervention. Visualization and analysis of these results are enhanced with Business Intelligence tools; we implement BI with Power BI to create interactive dashboards showing predictions, trends, and alerts generated by the models.
Process automation is another fundamental pillar. By combining MASHT with automated workflows, companies can respond to events in milliseconds. Our team at Q2BSTUDIO designs automation solutions that integrate time series classification with control systems, ERP, or IoT platforms. All of this is built on custom software that adapts to each client's specific needs, whether in healthcare, manufacturing, or services.
In summary, time series classification using random convolutional features combined with tabular foundation models represents a significant advance over traditional methods. MASHT demonstrates that it is possible to achieve state-of-the-art performance without task-specific training, facilitating its adoption in dynamic environments with limited resources. At Q2BSTUDIO we are ready to help organizations implement these technologies, from initial consulting to custom application development, cloud integration, cybersecurity, artificial intelligence, and data visualization. The future of time series classification is promising, and with the right approach, any company can leverage it to improve operations and make more informed decisions.
