Final Report Summary - HAROS-HD (Hybrid Adaptive Robust Optimization Strategy for EWIS High Dimensional systems)
Executive Summary:
In today’s engineering endeavor, it is common to run computer simulations to understand the behavior of complex systems and optimize their parameters to obtain satisfactory designs before actual physical prototypes are built. The objective of the engineer is not only the optimization of the systems but also to understand what makes a good design. The question - particularly in the early stage of the design process - is often not about finding the best parameter values, but is about what parameter ranges would generate competitive designs or solutions. In a more pragmatic simulation level, engineers often want to confine the simulation runs to parameter settings for which results are trustworthy. Such information may not be available until one actually runs the simulation (e.g. whether it crashes/converges or not).
The Haros-HD concept originated from engineering design situations in which accuracy of optimized result is as important as the efficient identification of the “good input space” (also called “feasibility region”).
Based on this consideration, the HAROS-HD challenge has been addressed by developing a different approach based on the pre-conditioning of the optimization problem with machine learning algorithms applied to engineering cases. As such, the effort is shifted from the optimization challenge to the ‘feature discovery’ process, where engineering features of the design and solution spaces are ‘discovered’ and exploited to perform a much faster and tailored optimization process.
This machine learning process starts with an analysis of the entities and relations of the electrical wire harness topology to identify subsystems that are independent or loosely coupled between themselves and that can be optimized separately with relatively small or no error (in case of total independence). Once the problem is decomposed according to its inherent structure, two machine learning algorithms are started to reduce the newly created subsystems and to learn, for each of them, the feasible region – that is the region of the output space where no constraints are active. In this context, the SOMBAS and the Deep Learning algorithms are used.
The output of this process is not only the definition of the feasible region, but also a set of optimization starting points that are best candidates for the subsequent subsystem optimization. The information discovered by the machine learning approach is used by specific optimization algorithms like Cross-Entropy and Annealed Hook&Jeeves that exploit this knowledge by starting from the best points found in the feasible region, thus removing the challenge to handle constraints (that, for the 48 harness, can be up to more than 10000). These optimization algorithms will then iterate till final convergence to the optimal solution for each subsystem. Once this is achieved, the overall system is then reassembled to take into account the various subsystem dependencies and converge to the final optimal configuration for the entire wire harness problem at hand.
The Haros-HD optimization strategy has been implemented and tested in a fully functional software prototype based on the Noesis Optimus 10.x platform and will be made available, together with its key enabling technologies, as commercial modules of the platform itself.
Project Context and Objectives:
HAROS-HD project general aim is to develop an innovative large scale optimization methodology that will help to address, manage and solve the high complexity of EWIS optimization problem.
As such, solving the problem of complexity might be one of the major goals of the 21st century: in the engineering field this is translated into the high number of parameters to take into consideration, as development projects are getting bigger and often too complex to be grasped entirely by one single person.
In particular, complexity in modern aircrafts is increasing significantly. They incorporate more electric systems than before since other subsystems that used to be pneumatic or hydraulic are being replaced by electric ones. As a consequence, the wire harnesses that are used to connect those systems to each other must also convey more signals and power.
Electrical systems are used for flight control, sensors, engine control, flight management, communication, in-flight entertainment and many more systems. Connecting the electrical power sources and consumers throughout the aircraft is done by the Electrical Wiring Interconnection System (EWIS). The EWIS is the entire collection of electrical wiring (conductors), connectors, bus bars, shielding, sleeves, pressure seals, brackets, etc... in the aircraft.
The entire wiring system is produced in modular components (for manufacturing and production purposes), so-called wire harnesses in which wires are bundled. At the final assembly those wire harnesses are integrated by attaching all the connectors.
For manufacturing and assembly reasons, production breaks are used. The electrical cables linking sources and loads pass through several harnesses. These cables are sized for the current they have to carry, with respect to some thermal and voltage drop constraints. As such, the cable gauge sizing problem is a multi-physics problem. An electrical link between a source and a load can have several cables with different gauges, allowing mass optimisation. Thermal and electrical aspects are to be taken into account in the first place, but other type of constraints exist (e.g. connectors’ properties, electromagnetic environment created by the current return network, the structure of the aircraft, the fuselage, and the couplings that may occur between all these elements) thus increasing the number of total variables and constraints of the design problem. Given the large amount of design variables and input/output constraints that describe the design space of real gauge sizing problems, the challenge of optimizing such system is considered to be a high-dimensional, non-linear and discrete/continuous problem.
This project has researched and developed a number of specific technologies and methods to address the challenge of optimizing complex and high dimensional systems. The use of unique machine learning techniques, implicit metamodeling using self-organizing maps and graph decomposition techniques has contributed to achieve a flexible, adaptive and scalable optimization strategy that could solve the EWIS problem. The resulting hybrid, adaptive and robust optimization strategy has demonstrated the optimization of high dimensional systems (HAROS-HD, Hybrid Adaptive Robust Optimization Strategy for High Dimensional systems).
Project Results:
The main tangible results of the Haros-HD project are the key enabling technologies and the high dimensional optimization strategy. These technologies and the optimization strategy have been implemented and deployed in a fully functional software prototype.
More specifically, a number of new technologies have been researched and implemented, starting from the current state of the art. These can be summarized as follows:
• Advanced graph decomposition techniques, aimed at structuring the information related to the EWIS problem and extracting dependencies and useful information for problem decomposition and sub-structuring.
• Feasible region identification algorithm (named Deep Learning), capable of sampling the design space and learning about the feasible region, given the current constraints. This algorithm, acting as an advanced classifier, is coupled with a special type of Neural network response model that, when coupled with the deep learning algorithm, is capable of creating an advanced surrogate model of the EWIS discrete system.
• Optimization based self organizing maps (SOMBAS) algorithm that, together with the Deep Learning approach, allow for quick space filling of the feasible region and identify the feasible boundaries, even if they are disconnected regions. The SOMBAS algorithm is a space filling approach based on the results of the Deep Learning one and produces information that is used in the subsequent steps. Together with SOMBAS also Interaction Indexes have been developed to extract information about the correlation between inputs and outputs with a local approach (i.e. not averaged over the whole domain but locally in each subregion of the domain).
• Adaptive DOE techniques that speed up the execution of both SOMBAS and Deep Learning, in order to minimize the amount of simulations to be performed.
• Advanced Cross Entropy and Hook&Jeeves algorithms. These have been largely enhanced to handle very complex and large scale problems and sub-problems and are particularly efficient in discrete problems. Cross Entropy uses a probabilistic approach to convert discrete problems into continuous ones, while the modified Hook and Jeeves explores the design space more efficiently, thanks to SOMBAS, Deep Learning, Graph decomposition and Cross Entropy results altogether.
Thanks to these technologies, it has been possible to design and implement a complete hybrid optimization procedure capable of handling high dimensional and large scale optimization problems. The final HAROS-HD prototype has been used to perform EWIS optimization on a 48 harness cable sizing problem with more than 400 design variables and 10000 constraints.
Potential Impact:
The main results of the project in terms of technologies and optimization strategy have led to the creation of a software prototype for a professional solution to the burdensome and time consuming problem of optimizing the Electrical Wiring Interconnection System. This software prototype has been developed based on the Noesis Optimus 10.x process integration and design optimization platform with the goal to achieve the reduction of global weight for an aircraft wire harness while ensuring a reliable and safe sizing for the strategic device that distributes electrical power all over the aircraft.
The competitiveness of the European providers for Electrical harness (based on inputs received by the Topic Manager) will increase by the drastic time reduction of this critical design phase. Also Fokker Elmo has raised interest in the platform during the final dissemination event.
As such, the expected impacts of the project are:
• Reduced design time for the EWIS optimization problem
• Extendability: can be applied to all optimization problems. HAROS-HD will learn about the design space and employ the appropriate algorithms as it proceeds toward finding an optimized solution independently of the number and type of design variables and constraints.
• Flexibility: HAROS-HD will intelligently adapt the optimization strategy by selecting the most appropriate method to use
• Scalability: It can be used for high dimensional problems
• application of advanced machine learning (ML) tools, based on surrogate models, to increase the overall efficiency of computer aided engineering processes,
This powerful optimization strategy will be industrialized and commercialized as part of the Noesis Optimus product family.
In order to promote these customer benefits, a final dissemination event has been organized and held on September 21st, 2015 to raise awareness of the project results to the scientific community and relevant industry. This event has been hosted by Safran Tech in Paris and 25 persons attended (including persons from Safran, Turbomeca, Thales, Fokker and others).
List of Websites:
No public website is available.
Relevant contact details are:
Roberto d'Ippolito - roberto.dippolito@noesissolutions.com
In today’s engineering endeavor, it is common to run computer simulations to understand the behavior of complex systems and optimize their parameters to obtain satisfactory designs before actual physical prototypes are built. The objective of the engineer is not only the optimization of the systems but also to understand what makes a good design. The question - particularly in the early stage of the design process - is often not about finding the best parameter values, but is about what parameter ranges would generate competitive designs or solutions. In a more pragmatic simulation level, engineers often want to confine the simulation runs to parameter settings for which results are trustworthy. Such information may not be available until one actually runs the simulation (e.g. whether it crashes/converges or not).
The Haros-HD concept originated from engineering design situations in which accuracy of optimized result is as important as the efficient identification of the “good input space” (also called “feasibility region”).
Based on this consideration, the HAROS-HD challenge has been addressed by developing a different approach based on the pre-conditioning of the optimization problem with machine learning algorithms applied to engineering cases. As such, the effort is shifted from the optimization challenge to the ‘feature discovery’ process, where engineering features of the design and solution spaces are ‘discovered’ and exploited to perform a much faster and tailored optimization process.
This machine learning process starts with an analysis of the entities and relations of the electrical wire harness topology to identify subsystems that are independent or loosely coupled between themselves and that can be optimized separately with relatively small or no error (in case of total independence). Once the problem is decomposed according to its inherent structure, two machine learning algorithms are started to reduce the newly created subsystems and to learn, for each of them, the feasible region – that is the region of the output space where no constraints are active. In this context, the SOMBAS and the Deep Learning algorithms are used.
The output of this process is not only the definition of the feasible region, but also a set of optimization starting points that are best candidates for the subsequent subsystem optimization. The information discovered by the machine learning approach is used by specific optimization algorithms like Cross-Entropy and Annealed Hook&Jeeves that exploit this knowledge by starting from the best points found in the feasible region, thus removing the challenge to handle constraints (that, for the 48 harness, can be up to more than 10000). These optimization algorithms will then iterate till final convergence to the optimal solution for each subsystem. Once this is achieved, the overall system is then reassembled to take into account the various subsystem dependencies and converge to the final optimal configuration for the entire wire harness problem at hand.
The Haros-HD optimization strategy has been implemented and tested in a fully functional software prototype based on the Noesis Optimus 10.x platform and will be made available, together with its key enabling technologies, as commercial modules of the platform itself.
Project Context and Objectives:
HAROS-HD project general aim is to develop an innovative large scale optimization methodology that will help to address, manage and solve the high complexity of EWIS optimization problem.
As such, solving the problem of complexity might be one of the major goals of the 21st century: in the engineering field this is translated into the high number of parameters to take into consideration, as development projects are getting bigger and often too complex to be grasped entirely by one single person.
In particular, complexity in modern aircrafts is increasing significantly. They incorporate more electric systems than before since other subsystems that used to be pneumatic or hydraulic are being replaced by electric ones. As a consequence, the wire harnesses that are used to connect those systems to each other must also convey more signals and power.
Electrical systems are used for flight control, sensors, engine control, flight management, communication, in-flight entertainment and many more systems. Connecting the electrical power sources and consumers throughout the aircraft is done by the Electrical Wiring Interconnection System (EWIS). The EWIS is the entire collection of electrical wiring (conductors), connectors, bus bars, shielding, sleeves, pressure seals, brackets, etc... in the aircraft.
The entire wiring system is produced in modular components (for manufacturing and production purposes), so-called wire harnesses in which wires are bundled. At the final assembly those wire harnesses are integrated by attaching all the connectors.
For manufacturing and assembly reasons, production breaks are used. The electrical cables linking sources and loads pass through several harnesses. These cables are sized for the current they have to carry, with respect to some thermal and voltage drop constraints. As such, the cable gauge sizing problem is a multi-physics problem. An electrical link between a source and a load can have several cables with different gauges, allowing mass optimisation. Thermal and electrical aspects are to be taken into account in the first place, but other type of constraints exist (e.g. connectors’ properties, electromagnetic environment created by the current return network, the structure of the aircraft, the fuselage, and the couplings that may occur between all these elements) thus increasing the number of total variables and constraints of the design problem. Given the large amount of design variables and input/output constraints that describe the design space of real gauge sizing problems, the challenge of optimizing such system is considered to be a high-dimensional, non-linear and discrete/continuous problem.
This project has researched and developed a number of specific technologies and methods to address the challenge of optimizing complex and high dimensional systems. The use of unique machine learning techniques, implicit metamodeling using self-organizing maps and graph decomposition techniques has contributed to achieve a flexible, adaptive and scalable optimization strategy that could solve the EWIS problem. The resulting hybrid, adaptive and robust optimization strategy has demonstrated the optimization of high dimensional systems (HAROS-HD, Hybrid Adaptive Robust Optimization Strategy for High Dimensional systems).
Project Results:
The main tangible results of the Haros-HD project are the key enabling technologies and the high dimensional optimization strategy. These technologies and the optimization strategy have been implemented and deployed in a fully functional software prototype.
More specifically, a number of new technologies have been researched and implemented, starting from the current state of the art. These can be summarized as follows:
• Advanced graph decomposition techniques, aimed at structuring the information related to the EWIS problem and extracting dependencies and useful information for problem decomposition and sub-structuring.
• Feasible region identification algorithm (named Deep Learning), capable of sampling the design space and learning about the feasible region, given the current constraints. This algorithm, acting as an advanced classifier, is coupled with a special type of Neural network response model that, when coupled with the deep learning algorithm, is capable of creating an advanced surrogate model of the EWIS discrete system.
• Optimization based self organizing maps (SOMBAS) algorithm that, together with the Deep Learning approach, allow for quick space filling of the feasible region and identify the feasible boundaries, even if they are disconnected regions. The SOMBAS algorithm is a space filling approach based on the results of the Deep Learning one and produces information that is used in the subsequent steps. Together with SOMBAS also Interaction Indexes have been developed to extract information about the correlation between inputs and outputs with a local approach (i.e. not averaged over the whole domain but locally in each subregion of the domain).
• Adaptive DOE techniques that speed up the execution of both SOMBAS and Deep Learning, in order to minimize the amount of simulations to be performed.
• Advanced Cross Entropy and Hook&Jeeves algorithms. These have been largely enhanced to handle very complex and large scale problems and sub-problems and are particularly efficient in discrete problems. Cross Entropy uses a probabilistic approach to convert discrete problems into continuous ones, while the modified Hook and Jeeves explores the design space more efficiently, thanks to SOMBAS, Deep Learning, Graph decomposition and Cross Entropy results altogether.
Thanks to these technologies, it has been possible to design and implement a complete hybrid optimization procedure capable of handling high dimensional and large scale optimization problems. The final HAROS-HD prototype has been used to perform EWIS optimization on a 48 harness cable sizing problem with more than 400 design variables and 10000 constraints.
Potential Impact:
The main results of the project in terms of technologies and optimization strategy have led to the creation of a software prototype for a professional solution to the burdensome and time consuming problem of optimizing the Electrical Wiring Interconnection System. This software prototype has been developed based on the Noesis Optimus 10.x process integration and design optimization platform with the goal to achieve the reduction of global weight for an aircraft wire harness while ensuring a reliable and safe sizing for the strategic device that distributes electrical power all over the aircraft.
The competitiveness of the European providers for Electrical harness (based on inputs received by the Topic Manager) will increase by the drastic time reduction of this critical design phase. Also Fokker Elmo has raised interest in the platform during the final dissemination event.
As such, the expected impacts of the project are:
• Reduced design time for the EWIS optimization problem
• Extendability: can be applied to all optimization problems. HAROS-HD will learn about the design space and employ the appropriate algorithms as it proceeds toward finding an optimized solution independently of the number and type of design variables and constraints.
• Flexibility: HAROS-HD will intelligently adapt the optimization strategy by selecting the most appropriate method to use
• Scalability: It can be used for high dimensional problems
• application of advanced machine learning (ML) tools, based on surrogate models, to increase the overall efficiency of computer aided engineering processes,
This powerful optimization strategy will be industrialized and commercialized as part of the Noesis Optimus product family.
In order to promote these customer benefits, a final dissemination event has been organized and held on September 21st, 2015 to raise awareness of the project results to the scientific community and relevant industry. This event has been hosted by Safran Tech in Paris and 25 persons attended (including persons from Safran, Turbomeca, Thales, Fokker and others).
List of Websites:
No public website is available.
Relevant contact details are:
Roberto d'Ippolito - roberto.dippolito@noesissolutions.com
