ParticleTensorAI® – From Signal Sequences to Tensors

ParticleTensorAI® transforms time-resolved light-scattering data into structured input for AI analysis. Instead of evaluating only isolated values from individual particle events, the method can retain signal sequences from one or more detectors and arrange them as a multidimensional data structure called a Particle Tensor.

In this context, a tensor is a structured arrangement of numerical measurement data. It can preserve relationships between signal samples, detector channels, and the selected measurement interval. The term describes the data format used for computational analysis; it does not refer to a separate physical property of a particle.

The exact tensor format depends on the measurement task and the AI model. There is no single representation that must be used for every ParticleTensorAI® application.

The transformation from measured signals to a Particle Tensor can be understood in the following stages:

1. Record detector signals: optical detectors capture time-dependent light-scattering signals as droplets or particles pass through the measurement region.
2. Select a measurement sequence: a defined interval of the recorded data is retained for analysis. The interval can contain information from multiple particle interactions.
3. Arrange the data: signal samples are organised into a multidimensional structure. Depending on the chosen method, the representation can preserve timing information and relationships between detector channels.
4. Prepare the model input: the Particle Tensor is formatted consistently with the AI model developed for the measurement objective.
5. Evaluate the tensor: the model analyses the tensor and produces a classification or estimate defined for the task.

Any data processing steps must be defined consistently for both model development and later evaluation. The tensor representation should preserve the signal information needed for the analysis objective.

A signal sequence contains more than its maximum or average intensity. Its time-dependent structure can include changes in amplitude, waveform shape, timing, and relationships between signals from different detectors.

Representing the data as a Particle Tensor allows an AI model to evaluate these features together. This is useful when the analysis may depend on patterns across a measurement interval rather than on a single value or a separately identified particle event.

One possible way to represent signal data is as an image. In a published study of a coal monoethylene glycol slurry spray, detector signals were mapped to RGB image channels. The resulting Particle Tensor representations were labelled with average coal-particle size and mass concentration and used to train convolutional neural networks for classification.

This image-based representation is one example of how signal sequences can be converted into model input. Other tensor formats may be appropriate for different detector arrangements, measurement data, or analysis objectives.

For supervised AI analysis, Particle Tensors are associated with reference information, such as known material conditions or measurement results. The model learns patterns that relate the tensor data to these references.

The training data should represent the conditions for which the model is intended to be used. Independent validation is needed to assess how reliably the model performs on data that were not used during training.

• The Particle Tensor is a data representation, not a direct measurement result.
• The tensor format and data preparation depend on the measurement task and selected AI model.
• Reference labels influence what a supervised model can learn to classify or estimate.
• Model outputs require validation under relevant material and process conditions.

The KITopen study “AI-assisted monitoring of coal particles within droplets in a coal monoethylene glycol slurry spray” describes an image-based Particle Tensor approach for analysing light-scattering data. The work investigates the classification of average coal-particle size and mass concentration for the tested slurry-spray conditions.

Read the study in the KITopen repository.