Self-configuring classification of discrete process variables states

Accurately classifying and tracking discrete time-series process states from one or more process variables (e.g., input/output sensors, actuators, etc) is a key data modelling challenge that addresses many critical use cases in Industrial and Manufacturing monitoring operations. Common use cases include identifying causal relationships between process variables to understand dependencies in an automated production line, or detecting when vibration sensors exhibit different behavioural patterns that indicate an early warning of component failure. The challenge can be formulated as a multidimensional time-series clustering task for univariate or multivariate time-series data, accounting for arbitrary sampling volume and rates, process states that emerge as complex time-dependent shapes in continuous process variables (e.g., sensor value as a univariate dataset) or mixed continuous and stateful process variables (e.g., sensor and actuator/program counter as a multivariate dataset), and state classification latency. By employing a novel temporal multidimensional feature matrix, a composite of time-orientated shapes may be provided as input to fit an unsupervised "state learner" model operating in Euclidean space. Combined with an optimiser for automatic determination of classification window and robust state identification, the unsupervised state learner model may be used to train a supervised time series classification model on univariate or multivariate process variables with highly scalable and accurate state classification in sub-millisecond latency.

View Research