Loading...
1 results
Search Results
Now showing 1 - 1 of 1
- Industrial Sensors Online Monitoring and Calibration Through Hidden Markov ModelsPublication . Martins, Alexandre Daniel Batista; Cardoso, António João Marques; Farinha, José Manuel TorresThis thesis aims to demonstrate a methodology able to diagnosis, through the Hidden Markov Model (HMM), the health state of production equipment, as well as the calibration state of sensors reading equipment. Through a well-defined methodology, the observations collected by the sensors are optimised to give input into a HMM, that are translated into hidden states, which represent the diagnosis of the equipment under study, being: State 1 - "Good working"; State 2 - "Warning"; State 3 - "Fault/Uncalibrated". After collecting the data, it goes through a cleaning process that will improve its quality and integrity. Then, a feature generation phase is performed. This phase is extremely important because the information can be managed for the desired equipment. It is through this stage that we can distinguish the diagnosis between the production equipment and the reading equipment. Next, a dimensional reduction of the data is performed, through Principal Component Analysis (PCA) and an extraction of new features that, although in smaller amounts, have more information each one. Then, the new data matrix is applied to a Clustering, performed by K-means, with the objective of grouping similar data within the same group. This will cause good working data to be in one cluster and bad working data to be in a different cluster. These clusters will be the optimized observable states that give input to the HMM. Subsequently, the HMM translates the observable states into a sequence of hidden states that represent the diagnosis of the equipment. Besides the methodology available to detect different types of information from the same data set, it has more capabilities, such as: imputing values in time series with few samples through Deep Neural Network (DNN) methods, namely the Multi-Layer Perceptron (MLP) model; performing the equipment health status prognosis through the Deep Neural Network (DNN), the Gated Recurrent Unit (GRU).