Thomas Rieger

Abstract

In industrial production environments, the maintenance of equipment has a decisive influence on costs and on the plannability of production capacities. In particular, unplanned failures during production times cause high costs, unplanned downtimes and possibly additional collateral damage. Predictive Maintenance starts here and tries to predict a possible failure and its cause so early that its prevention can be prepared and carried out in time. In order to be able to predict malfunctions and failures, the industrial plant with its characteristics, as well as wear and ageing processes, must be modelled. Such modelling can be done by replicating its physical properties. However, this is very complex and requires enormous expert knowledge about the plant and about wear and ageing processes of each individual component. Neural networks and machine learning make it possible to train such models using data and offer an alternative, especially when very complex and non-linear behaviour is evident. In order for models to make predictions, as much data as possible about the condition of a plant and its environment and production planning data is needed. In Industrial Internet of Things (IIoT) environments, the amount of available data is constantly increasing. Intelligent sensors and highly interconnected production facilities produce a steady stream of data. The sheer volume of data, but also the steady stream in which data is transmitted, place high demands on the data processing systems. If a participating system wants to perform live analyses on the incoming data streams, it must be able to process the incoming data at least as fast as the continuous data stream delivers it. If this is not the case, the system falls further and further behind in processing and thus in its analyses. This also applies to Predictive Maintenance systems, especially if they use complex and computationally intensive machine learning models. If sufficiently scalable hardware resources are available, this may not be a problem at first. However, if this is not the case or if the processing takes place on decentralised units with limited hardware resources (e.g. edge devices), the runtime behaviour and resource requirements of the type of neural network used can become an important criterion. This thesis addresses Predictive Maintenance systems in IIoT environments using neural networks and Deep Learning, where the runtime behaviour and the resource requirements are relevant. The question is whether it is possible to achieve better runtimes with similarly result quality using a new type of neural network. The focus is on reducing the complexity of the network and improving its parallelisability. Inspired by projects in which complexity was distributed to less complex neural subnetworks by upstream measures, two hypotheses presented in this thesis emerged: a) the distribution of complexity into simpler subnetworks leads to faster processing overall, despite the overhead this creates, and b) if a neural cell has a deeper internal structure, this leads to a less complex network. Within the framework of a qualitative study, an overall impression of Predictive Maintenance applications in IIoT environments using neural networks was developed. Based on the findings, a novel model layout was developed named Sliced Long Short-Term Memory Neural Network (SlicedLSTM). The SlicedLSTM implements the assumptions made in the aforementioned hypotheses in its inner model architecture. Within the framework of a quantitative study, the runtime behaviour of the SlicedLSTM was compared with that of a reference model in the form of laboratory tests. The study uses synthetically generated data from a NASA project to predict failures of modules of aircraft gas turbines. The dataset contains 1,414 multivariate time series with 104,897 samples of test data and 160,360 samples of training data. As a result, it could be proven for the specific application and the data used that the SlicedLSTM delivers faster processing times with similar result accuracy and thus clearly outperforms the reference model in this respect. The hypotheses about the influence of complexity in the internal structure of the neuronal cells were confirmed by the study carried out in the context of this thesis.

Keywords

Artificial Neural Networks, Deep Learning, Performance aspects of Artificial Neural Networks, Predictive Maintenance, Internet of Things, Long Short-Term Memory Neural Network, Sliced Long Short-Term Memory Neural Network, Complexity in Artificial Neural Networks

Document Type

Thesis

Publication Date

2023

Download

Share

COinS