Content uploaded by Eduard Shevtshenko
Author content
All content in this area was uploaded by Eduard Shevtshenko on Jan 27, 2016
Content may be subject to copyright.
Intelligent decision support system as a tool for decision support in collaborative network of SME
Abstract
The intelligent decision support system for the network of collaborative enterprises will provide decision support for the network participants and is based on the use of a strategic (aggregate) planning model for creating the manufacturing management decisions. Paper includes a multi-agent model with the processes that the IDSS will be able to optimise in the network of collaborative enterprises. For these purpose IDEF methodology will be used. The computer network-centric multi-agent approach is intended to facilitate interactions between many agents participating in the product and manufacturing process development. Also the security aspects will be included, as important condition for successful collaborative work. In practical part the production plan optimisation computational example will be included with analysis of results.
The navigation capabilities of existing off-road unmanned ground vehicles are severely limited by the perception capabilities of the vehicles’ on-board sensors. The reliable perception distance of the on-board sensors does not surpass 40m limiting the driving capabilities of an UGV to those of a human driver in dense fog. The objective of this thesis is to generate ad hoc navigation maps for an UGV, enabling smarter path planning, faster movement and reduced energy consumption.
The thesis proposes an intelligent long-range navigation method that learns from the UGV local navigation system and extrapolates the knowledge about local environment into wider area using overhead imagery. The proposed navigation method has built-in failsafe allowing UGV to fall back to local navigation system when the long-range system is not adequate. The method has energy efficient implementation that utilizes heterogeneous hardware, enabling deployment on battery powered vehicles.
The thesis consists of five parts: introduction, short review of literature, theoretical foundations, practical experiments and conclusion. The introduction presents the objectives of the thesis and describes its structure. The following review of the literature chapter discusses problems in existing
off-road capable navigation systems and proposed solutions. In addition the review of the literature chapter offers a brief summary of works in areas related to this thesis – aerial imagery classification, path planning and combining of maps from multiple sources.
The review of the literature is followed by a theoretical section, that focuses on orthophoto analysis, cost map generation, path planning and brings forth a possible usage scenario for the results of this work. For orthophoto analysis we proposed convolutional neural networks based classifier, that
combines trainable feature extractors and a linear classifier. The useful properties of this classifier are described in detail in the context of off-road navigation task.
The classifier output is further analyzed in Bayesian framework to find confidence information attached to extracted feature vectors. The classifier outputs a feature for each input pattern, each element of this vector represents the likelihood that corresponding feature is present on input pattern. In order to account this likelihood into cost map generation we need to convert it to
probability. The conversion is done by extracting probability density functions for each feature class from classifier training set and using them in Bayes’ theorem.
For cost map generation we first define a cost of unknown terrain, it is defined to be lower than the cost of obstacles but higher than the cost of clearances. This constant is used if classifier uncertainty is zero, for nonzero values the cost will gravitate towards it as confidence decreases. In addition
we attach a weight to each feature class; the traveling cost of an area is weighted sum of feature probabilities that is offset by the cost of unknown terrain. A special care is taken to ensure the traveling cost of an area remains positive at all times, otherwise it would break path planning algorithms.
The produced cost map is used for path planning. The path planner itself is not the subject of this thesis, but significant performance advantages can be gained from combing path planner, cost map generator and terrain classifier.
The terrain classification step is a relatively expensive one performance wise, reversing the evaluation pipeline and triggering the classifier on need-to-know basis both increases the reaction time of the navigation system and reduces the energy consumption.
The theoretical section is followed by a practical one that focuses on evaluating the capabilities of proposed navigation system. The main objective of the theoretical section is to demonstrate the navigation systems path planning ability on a known terrain and its self-assessment ability on an
unknown terrain. The experiments made in the course of this chapter use manually labeled aerial imagery from Estonian Land Board database and satellite imagery form Google Maps database.
The classifier capability is evaluated by comparing the classifier output against manually labeled aerial and satellite imagery and against labels from a GIS database. The classifier has shown an excellent classification ability achieving over 90% correct classification rate on all the tests. For validating the Bayesian filter a set of weakened classifiers were prepared that were
trained with insufficient and noisy data. The experiments show that the processed classification result of those weakened classifiers has low confidence attached to features. The confidence information is used by cost map generation algorithm, which gravitates the cost of areas with low
confidence vectors towards the cost of unknown terrain.
The classifier capability analysis is followed by a set of combined tests that cover aerial imagery classification, cost map generation and path planning. The objectives of combined tests are twofold: to demonstrate the path planning capability of the navigation system as a whole and to show the
behavior of the system with weakened classifier. The combined tests show that the navigation system is able to work under bad conditions when the surroundings of UGV are unfamiliar to the classifier.
The practical section is concluded by a subsection that addresses the main weakness of the classifier – the convolutional neural networks are large systems that require equally large computing capacities to operate. Execution of the proposed classifier on a CPU is unreasonable, because it requires the full power of a multicore processor for real time operation, which is a burden
for a battery powered vehicle. The thesis proposes a solution executing on heterogeneous computing architecture, which reduces the energy consumption by more than 250 times and reduces analyzing time by more than 850 times.
The body of the doctoral thesis is summarized by a conclusion chapter, which outlines the main achievements and future-oriented ideas.
To be competitive and successful on the market place today and satisfy customers, companies have to make strenuous efforts, for example, to improve various Key Performance Indicators (KPIs). However, the most widely spread problem faced by many manufacturing companies is that companies know the problems they have, for instance, unreliable production processes, poor product quality, financial losses, delay in product delivery, but frequently, they are unaware of the root causes of these problems. To solve these problems, companies are trying to implement various quality improvement programs, tools and methodologies. In many cases, these measures are used quite successfully to attain the goals appointed, but at the same time, they are working separately from each other and as a result, are not effective and consolidated enough.
Therefore, in this research a framework that integrates various quality improvement tools and methodologies has been created. This framework enables continuous enhancement of the reliability of a production process that influences the improvement of product quality, cost and delivery, with lower expenditures by collecting production data (problems, failures) about production processes. Further, these data help to define the most critical operations in the process and improve them. As a result, the process variability is reduced and process reliability is increased, which in turn decreases product scrap and rework. A reliable production process can save resources (labour, time, money), consequently can provide better product quality, save money and reduce delivery time, which improves company revenue and customer satisfaction.
Finally, the presented new framework will be adapted into the database – Data Mart, which will play the role of a “dashboard” which allows monitoring production processes (collect data about production problems, failures), measuring and analysis (based on various charts) based on data for the previous day. In addition, the new Data Mart will be applied into the Information Systems (IS) environment with various tools (Product Data Management (PDM), Extract Transfer Load (ETL), Enterprise Resource Planning (ERP) system) that enable us to process different data from one system to another and derive new knowledge useful for business processes management, decision making and customer satisfaction.
By integrating various quality improvement tools and methodologies with Information Technologies (IT) and IS tools, it will hold a vital and successful role in company’s business. This kind of integration helps businesses improve the efficiency and effectiveness of their production and business processes, production team support and collaboration and managerial decision-making, for example, which KPI is more important for a product, a company and a customer. This new framework strengthens company competitive positions, enables the company to be more adaptable to the rapidly changing marketplace and increase financial revenue.
ResearchGate has not been able to resolve any references for this publication.