Extreme Learning Machines with frequency based noise filtering for prediction of critical digressions in a noisy industrial process

Many systems that are continuous in time are also susceptible to adverse digressions that can result in severe losses. Conventional algorithms tend to use signals from sensors and detect faults after incipience, consequent to which corrective measures can be taken for mitigation. From an alternate perspective, it may be visualized that when some of the state parameters of the system combine to form specific relationships, such adverse digressions are engendered. Therefore, identifying these combinatorial relationships immediately upon formation can help in mitigation of such adverse effects. The advent of IoT in industry helped bring real time data to computing platforms where Machine Learning techniques can potentially be developed to detect formation of such relationships. Traditional Neural Networks (NN) can be used as a Machine Learning tool, but they require lot of time for training, and moreover, because the industrial data comes with a significant noise component, further processing of the NN outputs are needed before they can be used for such prediction purposes. Extreme Learning Machines (ELM) require about two orders of magnitude less training time, but are even more susceptible to noise corruption than NN due to reasons discussed in this paper. In this work, a method is developed to first filter out noise from industrial sensory data in real time from its spectral content, and then its downstream use in ELMs designed in ensemble pattern is found to generate outputs for predicting adverse digressions that are at least as accurate as traditional neural networks but with significantly reduced training time.