Assembly systems for configuration and operational control

Abstract

A computer-implemented method for operating an assembly configuration system, the method comprising: receiving an input demand; receiving one or more datasets including data points each including data of the performance indicators and data of the therewith associated inputs, outputs, and sets of parameters; determining from the one or more datasets a combined set of values for the sets of parameters based on the input demand and on a desired combination of the performance indicators; configuring the sets of parameters by assigning the combined set of values thereto; and generating an execution plan as the second output based on the input demand and the configured sets of parameters A computer-implemented method for operating a reconfiguration assembly system.

Description

Assembly Systems for Configuration and Operational Control Field Of The Invention

[0001] The present invention relates to operating an assembly configuration system (ACS).The present invention further relates to operating a reconfiguration assembly system (RAS) including the ACS. Background Art

[0002] A Reconfigurable Manufacturing System (RMS) or a Reconfigurable AssemblySystem (RAS) is a system that is designed at the outset for rapid change in structure, as well asin hardware and software components, in order to quickly adjust product capacity for differentproducts and their variants in response to sudden changes in market or in regulatoryrequirements. The goal of such systems is to create a highly flexible manufacturing or assemblyenvironment that can be rapidly reconfigured to produce different products or components.Such systems have received increasing attention due to their ability to adapt to changing production needs and thereby improve efficiency and reduce costs. However, the optimal configuration of such systems can be difficult to determine, particularly given the complexity of the production processes involved.

[0003] The typical decision flow to plan, schedule and execute the production of a list of ordersconsists out of an offline optimization process where different solvers tackle a certain subpartof the full integrated problem. A rough configuration design or process plan is calculated usingdeterministic optimization and then scheduled for an order batch towards the specific machines. This results in a final execution plan which is then sent to an Assembly Execution System(AES) to run on the physical hardware to process the given order list. This process is repeatedfor each batch of orders.

[0004] It is known that through the deterministic, sequential analysis and optimization of theRMS, the produced execution plans remain rough estimates that need further analysis and better verification. Often, no waiting times are considered and infinite buffer sizes between stations are assumed in order to keep solutions tractable. These assumptions overestimate the throughput of the manufacturing systems and may cause issues during real execution or even during handover between different solvers. When these issues occur, either the system runs suboptimal or with degrading performance using the pre-calculated execution plans or a manual intervention is needed to recalculate plans based on the changed context conditions.

[0005] A challenging task in the operation of a RMS is process planning and scheduling.

[0006] Koren et al. [Koren, Y., Gu, X., Guo, W. (2018). Reconfigurable manufacturing systems:Principles, design, and future trends. In Frontiers of Mechanical Engineering (Vol. 13, Issue2, pp. 121–136). Higher Education Press.] discusses several solutions for process planning andline balancing of reconfigurable systems in order to better model and solve the added complexity of reconfigurable systems versus dedicated systems. Their state of art to maximize system productivity illustrates the sequential approach where the integrated problem of process planning and scheduling is still treated as two separate, independent problems. In a first stage, the first problem of process planning is tackled as a macro-level process that determines thesequence of operations or processes needed to produce a part. For RMS or RAS, this is donethrough Variant Process Planning or variants. Plans for each part are generated based on optimizing a cost-based KPI (e.g. resource costs, hand-over time) and different meta-heuristic techniques have been explored. In a second stage, these process plans for an order batch are scheduled over the available reconfigurable machines using optimal scheduling or line balancing. This involves a second set of solver models that handle the scheduling on reconfigurable hardware based on performance-based KPI (e.g. makespan).

[0007] Yelles-Chaouche et al. [Yelles-Chaouche, A. R., Gurevsky, E., Brahimi, N., Dolgui, A.(2021). Reconfigurable manufacturing systems from an optimisation perspective: a focusedreview of literature. International Journal of Production Research, 59(21), 6400–6418] givesan overview of RMS from an optimization perspective. Issues occur because of this sequential approach to the integrated problem. The first stage of process planning solvers estimates independent plans for each assembly based on minimal cost. However, it does not have acomplete view on total makespan, as it lacks the detailed information when and on whichmachines the plan will be scheduled. This is calculated in the second stage and only then a makespan time can be estimated for an order batch.

[0008] Several other authors tackle the problem by formulating control frameworks thatcontrol the individual optimization solvers. They monitor KPI’s and allow to iterate over different instances of solvers with built in feedback control loops. Xu et al [L. D. Xu, C. Wang, Z. Bi and J. Yu, "AutoAssem: An Automated AssemblyPlanning System for Complex Products," in IEEE Transactions on Industrial Informatics, vol.8, no. 3, pp. 669-678, Aug. 2012, doi: 10.1109 / TII.2012.2188901.] proposes an automatedassembly planning system for complex products that integrates sequence and path planning, where sequence planning is equally split up in two optimization stages. The best plan is found based on a comparison of the results from those algorithms but the system still requires adesigner to manually explore and specify the parameters of the underlying (meta) heuristic algorithms. da Silva et al. [da Silva, R. M., Junqueira, F., Filho, D. J. S., & Miyagi, P. E. (2016). Control architecture and design method of reconfigurable manufacturing systems. ControlEngineering Practice, 49, 87–100] proposes a control architecture for RMS based on Petri netsand holons. It orchestrates fast reconfiguration of the hardware but it relies on the optimization processes as discussed above. Azab et al. [Azab, A., ElMaraghy, H., Nyhuis, P., Pachow-Frauenhofer, J., & Schmidt, M. (2013). Mechanics of change: A framework to reconfigure manufacturing systems. CIRP Journal of Manufacturing Science and Technology, 6(2), 110–119.https: / / doi.org / 10.1016 / j.cirpj.2012.12.002] discusses a framework for reconfigurable systemsthat introduces control loops that are triggered on system KPI’s to recalculate new plans orschedules. It is acknowledged in Azab et al. that reconfiguring a manufacturing system is acritical and lengthy process.

[0009] From an industry perspective, several data frameworks have been proposed that evolvebusiness intelligence (BI) targeted at tactical or strategic level business decisions towards theuse of BI to manage and optimize daily business operations, termed operational BI. ColinWhite [White C (2006) The next generation of business intelligence: operational BI. BIResearch, Ashland, OR] mainly focuses on data integration through standardization protocolsfor operational systems and operational BI reporting. Decision automation is a specific focus and currently centers on software agents that notify users on business issues and requirements that need urgent actions with the objective to reduce decision latency. Functionality can range from simple alerting to action recommendation or automated decision agents.

[0010] Rolling horizon planning is used in practice to reduce the complexity of the jointprocess planning problem, as well as the required CPU time to produce a solution. This at the expense of solution quality. In rolling horizon planning settings, the planning horizon is divided into decision periods, which effectively splits the planning problem into smaller subproblems.

[0011] Despite the huge efforts that have been made in this area, the operation of a RMS orRAS still present major challenges, such as suboptimal production, degrading performance, and requirement for manual intervention.

[0012] In order to overcome these limitations, specific requirements must be met for the RMSor RAS, such as reliability and efficiency in the planning and scheduling processes and possiblyin a distribution process.

[0013] Therefore, there is still a need for a method and system that can effectively andefficiently operate a RMS or RAS, specifically by controlling the RMS or RAS and optimizingproduct properties and production costs. Summary of the Disclosure

[0014] The present inventors have developed a method of running integrated processes of aplanning process and a scheduling process. Surprisingly, it was found that by simultaneouslytuning parameters in both the planning process and the scheduling process, a higher impact ofthese parameters could be achieved on the overall functioning of the system and itscomponents. Moreover, as will be shown in the appended examples, the inventors have furtherfound that the method can provide reduced makespan and improved on-time deliveryperformance of the system. This allows for more efficient operation of the RAS and its controlarchitecture to optimize product properties, production costs and production time, without theneed for more complex integrated process models.

[0015] In the context of this disclosure, the schedule defines the order in which tasks areexecuted and the detailed routings over machines. In the context of this disclosure, the processplan determines the resources, operations and optionally routings that are needed tomanufacture workpieces into parts or products. Routing is an optional aspect that can beestimated by planning, relevant in flexible assembly where the assembly path over or in the workstations needs to be determined in addition to the machine / people resources.

[0016] In the context of this disclosure, the term resources is intended to identify machines,tools, stock, space in the workshop floor and / or people.

[0017] In the context of this disclosure, the term input is intended to identify any unmodifiableattributes of an individual decision problem. They set the target of the decision problem, e.g., the number of units of a given product type that needs to be assembled or the sequence of jobs that need to be allocated.

[0018] In the context of this disclosure, the term parameter is intended to identify the way theproblem formulation is understood by the solver, i.e., they define the boundaries of the decision problem for the given target. This includes the max number and type of machines, operators and tools that can be used by the assembly line, the estimated weights for multi-criteriadecisions, etc. . Additionally, the term parameter is intended to influence how the solutionmechanism of the solver behaves. This includes parametrisation of metaheuristic algorithmsfor mathematical optimization, e.g. the population size of a genetic algorithm, etc . Eachparameter is assigned a value allowing the parameterized solver to generate a specific outputbased on an input, dependent of the values of the parameters.

[0019] In the context of this disclosure, the term KPI is intended to identify the critical (key)quantifiable indicators of progress or performance of an intended result. This can includeindividual indicators on manufacturing quality or assembly line performance, or can include aggregated indicators like Overall Equipment Effectiveness..

[0020] According to a first aspect of the present invention, there is provided a computer-implemented method for determining, from one or more datasets including data of the performance indicators and data of the therewith associated inputs, outputs, and sets of parameters, a combined set of values for a set of parameters wherein the parameters influence the output of a solver for a given input and dynamically control performance of planning and scheduling processes of an assembly configuration system, ACS, by generating one or more execution plans under the values for the sets of parameters, the computer-implemented method comprising the steps of determining a combined set of values for parameters to be reconfiguredfrom the one or more datasets by (i) determining a set of values for the parameters to bereconfigured for a first solver and a first output from a first dataset that is based on the given input, (ii) determining a partial set of configured parameters, and, for a desired performance indicator, mapping the first output from the first dataset to a second input of a second dataset to determine a second set of values for the parameters to be reconfigured for a second solver and mapping a second output from the second dataset based on (i) the second input, (ii) the partial set of configured parameters, and (iii) a desired performance indicator.

[0021] Advantageously, the method of the present invention ensures a combination ofimproved performance of the system and reduced production cost, especially when the execution plan is intended to be executed in a reconfigurable system where assembly lines have variability and changing conditions. It will be understood by a person skilled in the art that the planning and scheduling processes are intended to be implemented by means of control signals.

[0022] The first solver in the Assembly Configuration System (ACS) is responsible foroptimizing the planning process. This involves determining the resources, operations, and theirrouting needed to manufacture workpieces into parts or products. The first solver takes an inputdemand and a first set of configured parameters, and generates a first output, which is evaluated by a first performance indicator. Essentially, the first solver creates a process plan that outlines the sequence of operations and the resources required for production.

[0023] The second solver in the ACS is responsible for optimizing the scheduling process. Ittakes the first output generated by the first solver as its input and determines the order in whichtasks are executed and the detailed routings over the ACS. The second solver takes an inputthat corresponds to the output of the first solver and a second set of configured parameters, and generates a second output, which is evaluated by a second performance indicator. Essentially, the second solver creates a schedule that specifies the timing and sequence of tasks to be performed on the resources identified in the process plan, advantageously ensuring efficient execution and optimal performance.

[0024] The method of the invention uses one or more datasets that include data on performanceindicators (KPIs), inputs, outputs, and sets of parameters. These datasets help in determiningthe optimal values for the parameters that influence the output of the solvers. The method aimsto determine a combined set of values for the parameters that need to be reconfigured. These parameters control the performance of the planning and scheduling processes.

[0025] According to a second aspect of the present invention, the following further steps areadded to the first aspect of the invention: grouping data points in the one or more datasets wherein each group of data points is represented by a clustered data point that represents a distribution of input values, a distribution of output values, and a distribution of KPI’s that reflect the relation between input and output for a given parameter set, and wherein the clustered data point is stored as a clustered value for the input, a clustered value for the output, a KPI set and a parameter set.

[0026] Advantageously, these additional steps involve grouping data points in the datasets andrepresenting each group by a clustered data point. This clustered data point represents a distribution of input values, output values, and KPIs that reflect the relationship between input and output for a given parameter set. The clustered data point is stored as a clustered value forthe input, a clustered value for the output, a KPI set, and a parameter set. By grouping datapoints and using clustered data points, the method can simplify the datasets by reducing thenumber of individual data points, making it easier to manage and analyze the data, provide amore accurate representation (for each solver) of the relationships between inputs, outputs,parameters, and KPIs, leading to better optimization of the planning and scheduling processes, and speed up the process of determining the combined set of values for the parameters by using clustered data points, which represent groups of similar data points.

[0027] According to a further aspect of the present invention, a mapping step is added to themethod, mapping the first output of a clustered data point from the first dataset to a secondinput of a clustered data point from the second dataset. This mapping step further enhances theefficiency and accuracy of the optimization process by maintaining the relationship betweenthe first output and the second input, thereby ensuring that the optimization process isconsistent and reliable. The efficiency is enhanced by more effectively determining thecombined set of values for the parameters, which leads to a better performance of the planningand scheduling processes.

[0028] According to a further aspect of the present invention, the data set is enriched withpredicted data points generated by applying a machine learning model to a generated inputdemand. This step further enhances the optimization process by adding more data points to thedataset, providing a larger pool of information for the solvers to work with. The use of machinelearning to generate accurate predictions of outputs, and KPIs, from input demands, andparameters; advantageously leads to better optimization of the planning and schedulingprocesses. Moreover will it enhance the adaptability of the ACS by allowing the ACS to adaptto changing conditions and demands by using predicted data points to inform the optimization process.

[0029] According to yet another aspect of the present invention, a further method step involvesthe application of a machine learning model on combined properties of the planning andscheduling processes. By learning from combined properties, the machine learning model can better understand the relationships between planning and scheduling, leading to more accuratepredictions. Moreover, does it ensure that the machine learning model considers both planningand scheduling processes, leading to better integration and optimization of the ACS.

[0030] According to another aspect of the present invention, there is provided a computer-implemented method of running integrated processes of a planning process and a scheduling process, the method comprising the following steps of controlling the planning process by a first solver by reconfiguring a first set of parameters by determining, for a first problem instance, a closest clustered input and retrieving an associated first performance indicator andvalues for a first parameter set; retrieving, for the planning process by a first solver for a firstproblem instance and a first parameter set, a first clustered output by determining for a firstproblem instance a closest clustered input and retrieving the clustered output associated to thefirst clustered input and the first parameter set; controlling the scheduling process by a secondsolver by reconfiguring a second set of parameters by determining for a first problem instanceand the first parameter set and the first clustered output, a second closest clustered input andretrieving an associated second performance indicator and values for the second parameter set; wherein the first clustered output corresponds to a set of machines that need to be allocated to produce a given demand defined by the first problem instance; and the first clustered output isthen used as input, together with the first problem instance to retrieve the second parameter setof the second solver; and the running integrated processes of the planning process and thescheduling process is controlled by a combined parameter set that controls the output of the first solver and the second solver to produce output of a given demand.

[0031] The integrated processes of planning and scheduling are controlled by a combinedparameter set that controls the output of the first solver and the second solver to produce theoutput of a given demand. This method advantageously ensures a combination of improvedsystem performance and reduced production costs, especially when the execution plan is intended to be executed in a reconfigurable system where assembly lines have variability andchanging conditions. Additionally, the method allows estimating whether the determinedcombined set of values, and as a result, the generated execution plan, would result in an estimated KPI (or combination of estimated KPIs) that would be desired when the execution plan is intended to be executed by the RAS.

[0032] According to a further aspect of the present invention, a method for running integratedprocesses of a planning process and a scheduling process in an assembly configuration system(ACS) is proposed, enabling more efficient and accurate execution of production demands inreconfigurable assembly systems. This method is carried out by a computer and comprises acontrolling step of the planning process by a first solver (S1), a controlling step of the scheduling process by a second solver (S2) and a step for running the integrated processes of planning and scheduling, wherein these integrated processes are controlled by a combined parameter set (P1c, P1a, P2c, P2a) that controls the output of the first solver (S1) and the second solver (S2) to produce the output of a given demand. The controlling step of the planningprocess by a first solver (S1) comprises the steps of reconfiguring a first set of parameters bydetermining, for a first problem instance (I1), a closest clustered input (I1j) and retrieving an associated first performance indicator and values for a first parameter set (I1j, KPI1, P1c, P1a), and retrieving for the planning process by a first solver (S1) for a first problem instance (I1) and a first parameter set (P1a, P1c), a first clustered output (O1k) by determining for a first problem instance (I1) a closest clustered input (I1j) and retrieving the associated clusteredoutput (O1k) associated to the clustered input (I1j) and the first parameter set (P1a, P1c). Thecontrolling step of the scheduling process by a second solver (S2), comprises the steps ofreconfiguring a second set of parameters by determining for a first problem instance (I1) and the first parameter set (P1a, P1c) and their associated first clustered output (O1k), a second closest clustered input (I2m) and retrieving an associated second performance indicator and values for the second parameter set (I2m, KPI2, P2c, P2a). The first clustered output (O1k) corresponds to a set of machines that need to be allocated to produce a given demand defined by the first problem instance (I1); and the first clustered output (O1k) is then used as input,together with the first problem instance (I1) to retrieve the second parameter set of the second solver (S2).

[0033] This method advantageously ensures a combination of improved system performanceand reduced production costs, especially when the execution plan is intended to be executed in a reconfigurable system where assembly lines have variability and changing conditions.

[0034] According to a further aspect of the present invention, a planning decision space (I1,O1,KPI1) and a scheduling decision space (I2, O2,KPI2) are generated. This step involvescreating a decision space for the planning process, which includes the inputs (I1), the outputs(O1), and the associated performance (KPI1). The planning decision space allows for acomprehensive view of all possible planning scenarios and their corresponding outcomes andassociated performance. By generating this space, the method can better analyze and optimizethe planning process, considering various inputs and their potential outputs and associatedperformance. This contributes to more accurate and efficient planning decisions. Similarly,generating a scheduling decision space involves creating a decision space for the schedulingprocess, which includes the inputs (I2), the outputs (O2), and the associated performance(KPI2). This space provides a detailed view of all possible scheduling scenarios and theircorresponding outcomes and associated performance. By having this decision space, themethod can better analyze and optimize the scheduling process, considering various inputs andtheir potential outputs and associated performance. This leads to more precise and effectivescheduling decisions. By generating decision spaces for both planning and scheduling, themethod can advantageously perform more thorough and detailed optimization. It allows for the exploration of a wider range of scenarios and the identification of the best possible solutions for both planning and scheduling processes.

[0035] Hence, and according to afurther aspect of the present invention, there is provided acomputer-implemented method for operating an ACS, wherein operating the ACS includesrunning integrated processes of a planning process and a scheduling process, wherein theplanning process determines the resources, operations and their routing that are needed tomanufacture workpieces into parts or products and wherein the scheduling process determinesthe order in which tasks are executed and the detailed routings over the ACS. The planningprocess is optimized by a first solver which determines a first output from a first input andwhich is evaluated by a first performance indicator, wherein the scheduling process isoptimized by a second solver which determines a second output from a second input and which is evaluated by a second performance indicator, wherein the first solver associated with a first set of parameters and the second solver associated with a second set of parameters, wherein thesecond input is based on the first output, wherein the planning and scheduling processes are controlled by the first and second set of parameters, respectively, the method comprising: a. receiving one or more input demands as the first input and optionally a (partial) setof configured parameters; b. receiving one or more datasets each including data of the performance indicators anddata of the therewith associated inputs, outputs, and sets of parameters; c. determining from the one or more datasets a combined set of values for theparameters to be (re)configured based on the first input, optionally the set of configuredparameters and on a desired combination of the performance indicators;d. configuring or changing the set of configured parameters by assigning the combined set of values thereto; and e. generating one or more execution plans based on the first input and the (reconfiguredsets of parameters, said execution plan comprises the set of resources needed to produce the input demand(s) as the first output and the schedule(s) for that set of resources thatschedules the jobs over the resources as the second output.

[0036] Advantageously, the method of the present invention ensures a combination ofimproved performance of the system and reduced production cost, especially when theexecution plan is intended to be executed in a reconfigurable system where assembly lines havevariability and changing conditions. It will be understood by a person skilled in the art that theplanning and scheduling processes are intended to be implemented by means of control signals.

[0037] Furthermore, by using the method, the behavioral operation of all components isfactored in, and causes for reduced performance and / or deadlock of the RAS can be identified and acted upon immediately, i.e., with minimal interruption and with the best parameters toproduce a product under the changed conditions. That means considerable savings in outlayand quality of the manufacturing / production process of products, e.g. a considerable reductionof intervention time needed by calling on the expertise of the planning expert and scheduling expert to reconfigure the line and the solvers in case of production disruptions like rush orders, or reducing the risk of fast but suboptimal human mitigation decisions that are often greedy and do not look at the impact on all connected assembly lines on all aspects of performance and cost. Thus, production time and cost of the technical product is considerably reduced.

[0038] Furthermore, by using one or more datasets including the particular data points, themethod ensures that the system is operated as a data-based system that includes recorded performance (e.g., behavior) of reconfigurable systems, linked to the context of the offlinesolvers and defining the input conditions and RAS configurations. This allows for operationalcontrol and / or tactical level recommendation in order to improve standard expert or rule-based control of flexible assembly.

[0039] In embodiments according to the present invention, the method further comprises:f. simulating the generated execution plan to generate at least one estimatedperformance indicator; g. calculating a distance between the at least one estimated performance indicatorand at least one performance indicator from the desired combination of theperformance indicators; and h. when the calculated distance exceeds a predetermined threshold, repeating stepsc to h until it is determined that the calculated distance does not exceed thepredetermined threshold.

[0040] In preferred embodiments according to the present invention, the step of simulating(i.e., step f) generates a combination of estimated performance indicators, and the step of calculating a distance (i.e., step g) calculates a distance between the combination of estimated performance indicators and the desired combination of the performance indicators.

[0041] In some embodiments according to the invention, the step of simulating (i.e., step f)generates an estimated output, preferably an estimated first output and an estimated second output.

[0042] This allows the method to estimate whether the determined combined set of values, andas a result, the generated execution plan would result in an estimated KPI (or combination ofestimated KPIs) that would be desired when the execution plan is intended to be executed bythe RAS. Thus, by performing simulations, the behavioral performance of all components isfactored in, and resulting KPIs can be identified at an early stage, i.e. before the productionsystem is being reconfigured and set to produce again under the changed conditions.

[0043] Furthermore, the method steps have the effect of increasing the robustness of the ACSand the RAS, as well as running the optimization of the processing and scheduling processes. This applies to both the optimization problem and the reconfiguration assembly infrastructure. Every change in the assembly system leads to additional risks and efforts, both in terms of timing, production downtime and additional personnel and material expenses.

[0044] In embodiments according to the present invention, the step of determining from theone or more datasets a combined set of values for the sets of parameters (i.e., step c) includesmapping the input demand to one of the data points.

[0045] In embodiments according to the present invention, the one or more datasets comprise:a first dataset including data points each including data of the performance indicator describingthe performance of the planning decision process and data of the therewith associated input, output, and a set of parameters; and a second dataset including data points each including data of the performance indicator describing the performance of the scheduling decision process and data of the therewith associated input, output, and a set of parameters. In this case, the stepof determining a combined set of values for the parameters to be reconfigured from the datasetsmay comprises determining a set of values for the parameters to be reconfigured of the firstsolver and the first output from the first dataset based on the first input, the (partial) set ofconfigured parameters and on a desired performance indicator; mapping the first output fromthe first dataset to a second input from the second dataset; and determining a second set ofvalues for the parameters to be reconfigured of the second solver and the second output fromthe second dataset based on the second input, the (partial) set of configured parameters and ona desired performance indicator. By mapping the first output to one of the data points of thesecond dataset, the method ensures that the combined set of values for the parameters to bereconfigured is determined more effectively and accurately. The dataset needed to describe thefull integrated planning and scheduling process are kept small, by splitting up the total systemperformance indicators over separate datasets of sets of parameters and their performanceindicators of the individual decision processes that can be generated independently of each other but are linked through the mapping of the first output to the second input.

[0046] In preferred embodiments according to the present invention, the data points in thedatasets are grouped in groups of data points with similar input and output values, wherein eachgroups of data points is represented by a clustered data point that represents a distribution ofinput values, a distribution of output values, a distribution of values of KPIs associated withthe distribution of inputs and outputs, clustered value of the outputs and a set of parametersassociated with the distribution of inputs and outputs and KPI values. In this case, the combinedset of values for the sets of parameters are determined based on the clustered data points,wherein the mapping step comprises mapping the first output (O1) of a clustered data pointfrom the first dataset to a second input (I2) of a clustered data point from the second dataset.

[0047] By storing clustered data points associated with (historical) input demands as well asthe expected output and the set of parameters in the datasets and mapping the received inputdemand and optionally the set of configured parameters to one of the clustered data points, themethod allows for efficiently mapping the received input demand and optionally the (partial)set of configured parameters, and thus, speed up the overall process of determining a combinedset of values for the parameters to be reconfigured and generating an execution plan.

[0048] In embodiments according to the invention, the data points in the dataset comprisehistorical data points including observed data points. The historical data points may further comprise estimated data points and / or predicted data points.

[0049] In preferred embodiments according to the invention, the dataset is enriched withpredicted data points generated by applying a machine learning model to a generated inputdemand, wherein the machine learning model is learned on combined properties of the planningand the scheduling processes.

[0050] By generating input demands and applying the machine learning model to the generatedinput demands, the method allows to enrich the dataset by including more data points,particularly including data on the generated input demands and the therewith associated KPIs,inputs, outputs, and sets of parameters. Furthermore, by learning the machine learning modelon the combined properties, the resulting predicted data points are more accurate than machine learning models learned on properties of each process separately.

[0051] In embodiments according to the present invention, wherein the step of determining acombined set of values for the parameters to be reconfigured (i.e., step c) includes selecting theperformance indicators in the one or more datasets closest to the desired combination of theperformance indicators.

[0052] By calculating the distance (i.e., in step c and / or step g), the method ensures that thecombined set of values for the parameters to be reconfigured is more accurately determined,and thus, the execution plan is more effectively generated. This is even more advantageouswhen the estimated KPIs are stored in addition to the already existing KPIs associated with thereceived input demand. This further allows for a more reliable determination of the combinedset of values (i.e., in step c).

[0053] In embodiments according to the present invention, each of the KPIs in a clustered datapoint is stored as a measure of central tendency (e.g., mean / average, median or mode) of aplurality of KPIs associated with the input demand.

[0054] In embodiments according to the present invention, operating the ACS includes runninga plurality of planning processes in parallel, wherein the plurality of planning processes share a resource pool.

[0055] In preferred embodiments according to the present invention, operating the ACS in aparallel input-output structure, including running a plurality of scheduling processes in parallel, wherein the plurality of scheduling processes is equal in number to the plurality of planning processes, wherein each of the plurality of scheduling processes is integrated with therespective of the plurality of planning processes. For example, the output of each planningprocess is the input of their respective scheduling process. Advantageously, the method ensuresthat the ACS, and thereby the RAS, is more efficiently operated.

[0056] The parallel operation of the processes allows to identify the resources required for eachtask and determine the most efficient way to allocate these resources. This can help maximize resource utilization, minimize idle time, and minimize costs.

[0057] In embodiments according to the present invention, operating the ACS includes runninga distribution process optimized by a third solver which determines a third output from a thirdinput and which is evaluated by a third performance indicator, wherein the third solverassociated with a third set of parameters, wherein the distribution process is integrated with the planning process and the scheduling process such that the third input is the first output and the second input is the third output, wherein the distribution process is controlled by the third set of parameters.

[0058] By optimizing the distribution process in addition to the planning and schedulingprocesses, additional components are factored in, thus a combined set of values for the sets of parameters of all the processes can be even more effectively determined. This is particularly advantageous when including the distribution process in any one of parallel structures described herein, since the method ensures maximizing and efficiency and productivity of theRAS while minimizing the risk of errors or damage to products.

[0059] According to a second aspect of the present invention, there is provided an assemblyconfiguration system (ACS) comprising: a means for carrying out the method as described herein.

[0060] In embodiments according to the present invention, the ACS further comprises one ormore datasets communicatively coupled to said means. This ensures that the communicationtime between the said means and the one or more datasets is reduced. This is particularlyadvantageous when the one or more datasets increases in size.

[0061] According to a third aspect of the present invention, there is provided a computer-implemented method of operating a reconfiguration assembly system (RAS) including an assembly configuration system (ACS) and an assembly execution system (AES) the method comprising executing, by the AES, an execution plan generated by the ACS as described herein.

[0062] The main advantage is that the ACS provides control of the optimization solvers whichare offline and as a control interface between the offline optimization (i.e., via the solvers) andonline execution systems (i.e., physical systems). By performing the offline step(s) of theoperating ACS described, the method ensures that online steps can be more efficiently run, and thus, allowing the ACS and / or the RAS to be operated more efficiently as well.

[0063] In embodiments according to the present invention, the method further comprisingreceiving, by the ACS from the AES, an observed performance indicator after executing the generated execution plan; and storing, by the ACS, the observed performance indicator in theone or more datasets associated with the input demand.

[0064] Advantageously, the method allows the assessment during operation of the RAS underwhat parameter settings and at what performance and cost a new configuration would improvethe current settings, and ensure an improved adaptation of the configuration. The initialparameter tunings of the individual components often need revisiting as conditions change, andfor an individual component this is difficult to decide without knowledge of the systembehavior. Next to the operational level, the same dataset can be used to generaterecommendation on what resources to activate for a given assembly line in the planning phase(i.e. a tactical level decision support). This is even more advantageous when the observed KPIsare stored in addition to the already existing KPIs (e.g., estimated KPIs) associated with the received input demand. This further allows for a more reliable determination of the combined set of values (i.e., in step c).

[0065] According to a fourth aspect of the present invention, there is provided areconfiguration assembly system (RAS) comprising: an assembly configuration system (ACS) as described herein; and an assembly execution system (AES) configured to execute an execution plan generated by the ACS as described herein.

[0066] According to a fifth aspect of the present invention, there is provided a computerprogram comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method as described herein.

[0067] According to a fifth aspect of the present invention, there is provided a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method as described herein.

[0068] Additional and alternative objects of the present invention may be understood from thefollowing.Brief Description Of The Drawings

[0069] The present invention will be discussed in more detail below, with reference to theattached drawings, in which:

[0070] Figs. 1A-1B illustrate embodiments of a method according to the present invention;

[0071] Fig. 2 illustrates an assembly configuration system (ACS) comprised in areconfiguration assembly system (RAS); and

[0072] Figs. 3A-3C illustrate embodiments of a method according to the present invention.Description Of Embodiments

[0073] The following descriptions depict only example embodiments and are not consideredlimiting in scope. Any reference herein to the disclosure is not intended to restrict or limit the disclosure to exact features of any one or more of the exemplary embodiments disclosed in the present specification.

[0074] Furthermore, the terms first, second, third and the like in the description and in theclaims are used for distinguishing between similar elements and not necessarily for describing a sequential or chronological order. The terms are interchangeable under appropriate circumstances and the embodiments of the invention can operate in other sequences than described or illustrated herein.

[0075] Furthermore, the various embodiments, although referred to as “preferred” are to beconstrued as exemplary manners in which the invention may be implemented rather than as limiting the scope of the invention.

[0076] The terms “manufacture”, “assembly”, and “production” will be used interchangeablyherein to refer to the act or process resulting in a product.

[0077] The product is a physical product and may include one of a chemical product (e.g.,polymers, solvents, and fertilizers), a pharmaceutical and / or biotechnical product (e.g.,vaccines and enzymes), furniture, electronics, vehicle, appliance, toy, medical devices, clothing, etc. Other physical products will be known to the person skilled in the art.

[0078] The term “comprising”, used in the claims, should not be interpreted as being restrictedto the elements or steps listed thereafter; it does not exclude other elements or steps. It needs to be interpreted as specifying the presence of the stated features, integers, steps or components as referred to, but does not preclude the presence or addition of one or more other features, integers, steps or components, or groups thereof. Thus, the scope of the expression “a device comprising A and B” should not be limited to devices consisting only of components A and B,rather with respect to the present invention, the only enumerated components of the device are A and B, and further the claim should be interpreted as including equivalents of those components.

[0079] In reconfigurable manufacturing / assembly systems, a solver is a software tool that isused to automate a process, such as planning, scheduling and distribution processes.

[0080] A “planning” solver is used in a planning process to determine the optimal sequence oftasks required to manufacture or assemble a product. The solver can help determine the optimalsequence by considering factors such as the availability of resources, production capacity,production time and customer demand. The solver is associated with a first set of parametersincluding constraint parameters (e.g., available set of resources and / or resource constraints) and algorithmic parameters (e.g., in the case of using a Genetic Algorithm (GA): GA population size, crossover and mutation operations, etc.; in the case of a Linear Program (LP): objective function, decision variables, and constraints).

[0081] An input of the planning process solver may be at least one of: set of orders (i.e., theproduct type, quantity, delivery date, and any special requirements that the customer may have), product assembly sequences, production data (i.e., information on current and expected production levels, including demand forecasts, inventory levels, and production capacity), product information (i.e., details on the products being manufactured, including their specifications, volumes, and required production times), production costs. An output of theplanning process solver may be at least one of: production configuration (i.e., a detailed planfor the production process that outlines the required resources, production timelines, and expected output levels for each product), resource allocation plan, cost analysis (i.e., includingthe costs of raw materials, labor, and equipment), preferably production configuration. A keyperformance indicator of the planning process may be at least one of: total cost of ownership (TCO), capacity utilization, customer satisfaction, lead time, order accuracy, productionefficiency, material waste, and production cost, preferably TCO. An example of this first inputis a daily order including the number of demand units of each product that need to be produced, and the moment (time-period) of the day in which they need to be produced and delivered (a daily order is e.g. a 6x3 matrix where each (i,j) entry corresponds to the units of demand of thejth product for the ith time period of the day). An example of this first output is a set of resourcesand number of units of each resource that are to be allocated for the given day (also referred to as Resource pool), the layout the physical location of each resource instance available at each time period of the day, and the routing of all parts and products through the system (thesequence of tasks to be followed to assemble each product instance) for each time period in the day.

[0082] A “scheduling” solver is used in a scheduling process to assign specific times for eachtask to be performed. The solver can help generate schedules that minimize production time,maximize efficiency, and minimize the use of resources. The solver is associated with a secondset of parameters including constraint parameters (e.g., station cycle time, station sequence, buffer capacity, setup time, etc.) and algorithmic parameters (e.g., in the case of using a Genetic Algorithm (GA): GA population size, crossover and mutation operations, etc.; in the case of a Linear Program (LP): objective function, decision variables, and constraints).

[0083] An input of the scheduling process solver may be at least one of: production plan, orderinformation (e.g., details on customer orders, including the quantities, delivery dates, and shipping addresses), resource availability, production constraints (e.g., equipmentmaintenance, downtime or regulatory requirements). An output of the scheduling processsolver may be at least one of: job schedule (i.e., detailed schedule that outlines the sequence and timing of operations for the production process, including start and end times for each operation) and resource allocation plan, preferably job schedule. A key performance indicator(referred to herein as KPI) of the scheduling process may be at least one of: on-time delivery(OTD), makespan, production cycle time, downtime, setup time, changeover time, resourceallocation, production efficiency, and production output, preferably on-time delivery and / ormakespan. An example of this second input is the sequence of tasks to be scheduled, whichmay be the routing of all parts and products through the system (defined as part of the first output). An example of the second output is the start and end time of each task in the sequence of tasks for each time period in the day.

[0084] A “distribution” solver is used in a distribution process to determine the most efficientway to transport products and materials between different locations in the manufacturing / assembly line. The solver can help generate distribution plans that minimizetransportation time, reduce costs, and optimize the use of transport resources and inventory.The solver is associated with a third set of parameters including constraint parameters (e.g., delivery windows, lead times, transit times, availability of transportation resources, time- sensitive delivery requirements, distance between locations, the amount of materials to be transported, and the available transport vehicles) and algorithmic parameters (e.g., in the case of using a Genetic Algorithm (GA): GA population size, crossover and mutation operations, etc.; in the case of a Linear Program (LP): objective function, decision variables, and constraints).

[0085] In the planning process, resource pool changeover KPIs can be measured to assess theefficiency and effectiveness of resource allocation strategies. This involves evaluating the frequency and duration of changeovers, resource utilization levels, and the overall impact of resource reallocation decisions on production capacity and costs. Similarly, nervousness- related KPIs can be assessed during the planning process to evaluate the system's ability to respond to changes in demand or unexpected disruptions. This may include analyzing production lead time variability, the ability to adjust production schedules in response to demand fluctuations, and the accuracy of demand forecasting. In the scheduling process, these KPIs can be further monitored and evaluated. For example, changeover time KPIs can be tracked to optimize scheduling decisions and minimize idle time or disruptions caused by resource reconfigurations. Nervousness-related KPIs can help assess the responsiveness of the scheduling process to changing demand patterns and the ability to maintain a stable andefficient production flow. Overall, both planning and scheduling processes play a role inmeasuring these KPIs to improve resource utilization, production efficiency, and adaptabilitywithin manufacturing operations. Resource pool changeovers and nervousness can indirectlyimpact the distribution process by influencing production output, lead times, and overall supply chain efficiency.

[0086] Thus, constraint parameters refer to parameters that influence the way the problemformulation is understood by the solver. These can relate to, e.g., the maximum number ofresources of each type available at each moment in time, constraints on the type of tasks theycan perform, constraints on allocation or on reconfigurability capabilities, assembly line performance constraints, etc. The second set of parameters, the algorithmic parameters, refer to the set of parameters that influence the algorithms used to solve the planning problem. These could be, e.g., early stop parameters, optimality gap, pre-solve parameters or direction preference in branch and bound methods in the case of exact methods, mutation and crossover in the case of evolutionary-based (meta)heuristic algorithms, or probabilistic, stopping, or step size criteria in the case of search-based (meta)heuristic algorithms.

[0087] An input of the distribution process solver may be at least one of: finished products,order information, transportation and logistics data (e.g., information on available transportation options), inventory data (e.g., inventory levels). An output of the distributionprocess solver may be at least one of: delivery schedule (i.e., outlines the delivery dates andshipping methods for each sales order), inventory management plan (e.g., tracking and replenishing stock levels), logistics plan (e.g., transportation methods and routes for deliveringproducts). A key performance indicator of the planning process may be at least one of: on-time delivery, order accuracy, lead time, fill rate, cost per order, return rate, inventory accuracy.

[0088] A first aspect of the present invention will be described with reference to Fig. 1 andFigs. 3A-3B. Figs. 1A-1B and Figs. 3A-3B illustrate embodiments of a method for operatingan assembly configuration system (ACS) according to the present invention. Operating theACS includes running integrated processes of a planning process and a scheduling process.

[0089] As shown in Fig. 3A, the planning process is optimized by a first solver S1 whichdetermines a first output O1from a first input I1and which is evaluated by a first performance indicator KPI S1, wherein the scheduling process is optimized by a second solver S2which determines a second output O2 from a second input I2 and which is evaluated by a second performance indicator KPI S2, wherein the first solver S1is associated with a first set of constraint parameter(s) P1cand algorithmic parameter(s) P1aand the second solver S2is associated with a second set of constraint parameter(s) P2c and algorithmic parameter(s) P2a, wherein the second input I2 is based on the first output O1, wherein the planning and scheduling processes are controlled by a controller 313 controlling the first and second set of constraint and algorithmic parameters P1c, P1a, P2c, P2a.

[0090] The second input I2 is based on the first output O1, such as where the second input I2 isthe first output O1, or such as where the first output O1is processed and the processed first output is used as the second input I2, or such as when one or more additional processes, e.g., the distribution process described with reference to Fig. 3C, are included in between the planning and the scheduling processes and the output of the last of said one or more additional processes is used as the second input I2.

[0091] As will be understood from Figs. 3A-3C, operating the ACS includes running at leastthe planning and the scheduling processes, as described herein. One or more additional processes, such as the distribution process described herein, may be integrated with theplanning and scheduling processes, e.g., included in between and / or after the planning and thescheduling processes. For example, a first additional process having an input being the outputof the planning process and having an output being the input of the scheduling processes and a second additional process having an input being the output of the scheduling process.

[0092] As shown in Fig. 1A, the method according to the present invention comprises step a:receiving 101 an input demand as the first input I1; step b: receiving 102 one or more datasetsincluding data points each including data of the performance indicators KPI S1, KPI S2 anddata of the therewith associated inputs I1, I2, outputs O1, O2, and sets of parameters P1c, P1a, P2c,P2a; step c: determining 104 from the one or more datasets a combined set of values for the setsof parameters P1c, P1a, P2c, P2abased on the input demand and on a desired combination of the performance indicators KPI D1, KPI D2; step d: configuring 106 the sets of parameters P1c, P1a, P2c, P2a by assigning the combined set of values thereto; and step e: generating 108 an execution plan as the second output O2 based on the input demand and the configured sets of parameters P1c, P1a, P2c, P2a.

[0093] Step a (i.e., receiving 101 input demand) may be performed through a graphical userinterface (GUI), e.g., comprised in the ACS. However, the present invention is not limitedthereto. The input demand may be received by other communication means known in the art,e.g., from an external device by wireless communication.

[0094] Step b: (i.e., receiving 102 one or more datasets) may include retrieving the one or moredatasets comprised in the ACS. However, the present invention is not limited thereto. The oneor more datasets may be received from external systems, such as a cloud storage, an externalserver, etc.

[0095] Step c : (i.e., determining 104) may include comparing the input demand to the datapoints of the first input I1and comparing the desired combination of the performance indicators KPI D1, KPI D2 to the data of the performance indicators KPI S1, KPI S2 in the data points.Step c (i.e., determining 104) may further include calculating distances between theperformance indicators in the one or more datasets and the desired combination of theperformance indicators, wherein the combined set of values for the sets of parameters is determined to have the smallest distance from among the calculated distances.

[0096] In preferred embodiments according to the present invention, step c (i.e., determining104) includes mapping the input demand to one of the data points. This mapping may includecalculating distances between the input demand and data points of the first input I1, whereinthe data point to which the input demand is mapped is determined to have the smallest distancefrom among the calculated distances. Preferably, the data points of the first input I1 includeclustered data points of the first input I1, wherein the input demand is mapped to the clustereddata point.

[0097] Step d (i.e., configuring 106 the sets of parameters) may involve assigning thecombined set of values to the sets of parameters, thereby controlling the planning, scheduling processes and any other additional processes.

[0098] The generated execution plan is a result of the processes run in the ACS, which mayinclude at least information on equipment and resource allocation (i.e., which equipment and resources will be utilized to produce the desired products, and how they will be allocated acrossdifferent stages of production) and a production schedule (e.g., including the start and end timesfor each stage of production, and the expected completion times for each product). Theexecution plan may be transmitted to the assembly execution system (AES) directly or pre- processed before transmitting.

[0099] The step of transmitting 116 the execution plan for execution may be performedsubsequently after step 108 of generating 108 the execution plan. It is preferred that the methodfurther comprises performing the following steps f to h before step 116. Step f: simulating 110the generated execution plan to generate an estimated result, incl. at least one of an estimatedperformance indicator and estimated output of the respective process run in the ACS, such asa first estimated performance indicator KPI E1, a second estimated performance indicator KPIE2, or a combination of the first and second estimated performance indicators KPI E1, KPI E2;step g: calculating 112 a distance between the estimated performance indicator, such as the estimated first performance indicator KPI E1or the estimated second performance indicator KPI E2, and a performance indicator from the desired combination of the performanceindicators, such as the desired first performance indicator KPI D1 or the desired secondperformance indicator KPI D2; and step h: when the calculated distance exceeds 114 apredetermined threshold, repeating steps c to h (i.e., steps 104, 106, 108, 110, 112 and 114)until it is determined that the calculated distance does not exceed the predetermined threshold.

[0100] Step f (i.e., simulating 110) may include applying a mathematical model of theprocesses run in the ACS. The mathematical model may be learned on the one or more datasetsto calculate the behavior of the system as conditions evolve over time and / or as events occur.Step f (i.e., simulating 110) may include specifying the characteristics of the data, particularlythe input demand, the set of parameters and the solvers optimizing the processes in the ACS. Said data may be extracted from the generated execution plan.

[0101] The estimated result generated in step f (i.e., simulating 110) is preferably stored in theone or more datasets, wherein the estimated result is associated with the input demand and theassigned set of parameters on which the execution plan is based. It is preferred that at least theestimated KPI in the estimated result is stored in the one or more datasets.

[0102] The step(s) of calculating a distance in step c and / or the step g (i.e., determining 104 orcalculating 112 a distance) may include a distance metric known to the skilled person, such asEuclidean, Manhattan, Chebyshev or Jaccard. Calculating the distance in the feature spaceinvolves determining the dissimilarity or similarity between two data points, which arerepresented as vectors in the feature space and where each dimension of the vector correspondsto a specific feature. The feature space refers to the set of all features or attributes that are used to describe or represent a data point.

[0103] In step g (i.e., calculating 112 ), it is preferred that the distance is calculated betweenthe combination of estimated performance indicators of a plurality of processes run in the ACS,such as a combination of the estimated first and second performance indicators KPI E1, KPIE2, and the desired combination of the performance indicators, such as the combination of the desired first and second performance indicators KPI D1, KPI D2.

[0104] In step h (i.e., determining 114 whether the distance exceeds a predeterminedthreshold), the threshold may be determined by a user of the ACS, wherein the threshold provides a strictness to the difference / similarity between the estimated KPI(s) and the desired KPI(s). Thus, the lower the threshold the stricter the difference can be, i.e., the smaller the distance has to be. If the estimated KPI(s) lies outside the threshold, this triggers a search for new values for the set of parameters in the one or more datasets. Determining the new values for the set of parameters may be performed via an exhaustive search over the parameter space.However, other methods could be applied, such as local search or metaheuristic methods.Additional indicators may be introduced to make the decision more robust, e.g., total decisionlead time based on the individual service computational time and / or re-ranking of optimalsolutions based on robustness against future demand.

[0105] As shown in Fig. 1B, the method according to the present invention may be extendedin the case a full or partial set of preconfigured parameters is provided. In typical practice,parameters have been configured at start of process by planning and scheduling experts. A full configured set is then available and the manufacturing process can. If no disruptions or changes in the input or assembly context occur, the system can keep producing under this set of configured params as long as KPI fall within threshold. If disruption happens (e.g. increasedemand, machine failure, …) and a KPI exceeds threshold, a reconfigured set of parametersare estimated and simulated. If ok, the plan is transmitted for execution and system continues producing in its reconfigured parameter state.

[0106] As shown in Fig. 3C, operating the ACS may include running a distribution processoptimized by a third solver S3 which determines a third output O3 from a third input I3 and which is evaluated by a third performance indicator KPI S3, wherein the third solver S3is associated with a third set of constraint parameter(s) P3cand algorithmic parameter(s) P3a,wherein the distribution process is integrated with the planning process and the schedulingprocess such that the third input I3is the first output O1and the second input I2is the third output O3. Although not shown in Fig.3C, the planning, scheduling and distribution processes are controlled by a controller 313 controlling the first, second and third set of constraint and algorithmic parameters P1c, P1a, P2c, P2a, P3c, P3a.

[0107] The second input I2 may be the third output O3. However, it is may be that the secondinput I2is based on the third output O3, such as where the third output O3is processed and the processed third output is used as the second input I2, or such as when one or more additional processes, are included in between the distribution and scheduling processes and the output of the last of said one or more additional processes is used as the second input I2.

[0108] As shown in Fig. 3B, operating the ACS may be done in a parallel input structure,including running a plurality of planning processes in parallel and running a multi-input scheduling process integrated with the plurality of planning processes, such that each of the multi inputs is an output of a respective planning process. Fig. 3B shows a first and second planning processes are run in parallel, however more planning processes may be run in parallel.The first planning process is optimized by a first parallel solver S1,1 which determines a firstparallel output O1,1from a first parallel input I1,1and which is evaluated by a first parallel performance indicator KPI S1,1, wherein the first parallel solver S1,1 is associated with a first parallel set of constraint parameter(s) P1,1c and algorithmic parameter(s) P1,1a. A second planning process running in parallel to the first planning process is optimized by a second parallel solver S1,2 which determines a second parallel output O1,2 from a second parallel input I1,2 and which is evaluated by a second parallel performance indicator KPI S1,2, wherein the second parallel solver S1,2is associated with a second parallel set of parameters P1,2c, P1,2a.

[0109] The combination of performance indicators may be a linear combination (e.g., weightedsum) or a non-linear combination (e.g., quadratic sum).

[0110] The plurality of planning processes running in parallel may share a resource pool. Theresource pool may be managed by the controller 313 which monitors the availability and utilization of each resource in the resource pool. The controller 313 can automatically allocateresources to different manufacturing tasks by controlling the set of parameters of the any oneor more of the plurality of planning processes.

[0111] The ACS may be operated in a parallel input-output structure, including running aplurality of scheduling processes in parallel, wherein the plurality of scheduling processes is equal in number to the plurality of planning processes, wherein each of the plurality of scheduling processes is integrated with the respective of the plurality of planning processes, such that the output of each planning process is the input of the respective scheduling process. Advantageously, the method ensures that the ACS, and thereby the RAS, is more efficiently operated.

[0112] In alternative embodiments according to the present invention, operating the ACS in aparallel multi-input structure, including running a plurality of planning processes in paralleland running a multi-input scheduling process integrated with the plurality of planning processes, wherein the multi-input scheduling process is optimized by a second multi-input solver which determines the second output from a plurality of second inputs and which is evaluated by the second performance indicator. Having a multi-input scheduling process ensures that the behavioral performance of all the plurality of planning processes is factored in, and therefore, a combined set of values for the sets of parameters of the plurality of processes and the scheduling process can be more effectively determined. Advantageously, the method ensures that the ACS, and thereby the RAS is more effectively operated.

[0113] Fig. 3B shows a multi-input scheduling process which is optimized by a second solverS2 which determines a first output O2,1 and a second output O2,2 from a first input I2,1 and a second input I2,2and which is evaluated by a first performance indicator KPI S2,1and a second performance indicator KPI S2,2, wherein the second solver S2is associated with a second set of constraint and algorithmic parameters P2c, P2a. Although not shown, the second solver may alternatively determine a single output from the first input I2,1 and the second input I2,2 and may be evaluated by a single performance indicator.

[0114] The first input I2,1 is based on the first parallel output O1,1 and the second input I2,2 isbased on the second parallel output O1,2, such as where the first input I2,1 is the first parallel output O1,1and the second input I2,2is the second parallel output O1,2., or such as where the first parallel output O1,1 is processed and the processed first parallel output is used as the first input I2,1 and the second parallel output O1,2 is processed and the processed second parallel output is used as the second input I2,2, or such as when one or more additional processes, such as the distribution process described with reference to Fig.3C, are included in between each ofthe planning processes and the scheduling process and the output of the last of said one or moreadditional processes is used as the respective input of the scheduling process.

[0115] In preferred alternative embodiments according to the present invention, the ACS isoperated in a parallel mixed structure, including running a plurality of first planning processesand a plurality of first scheduling processes in parallel, and running a plurality of second planning processes in parallel and running one or more multi-input scheduling processes integrated with the plurality of second planning processes. Advantageously, the method ensures that the a first part (relating to the first planning and scheduling processes) of the ACS is more efficiently operated, and that the a second part (relating to the second planning and the one or more multi-input scheduling processes) of the ACS is more effectively operated.

[0116] As will be understood from Figs. 3A-3C, operating the ACS includes running at leastthe planning and the scheduling processes, as described herein. One or more additionalprocesses, such as the distribution process described herein, may be integrated with the planning and scheduling processes, e.g., included in between or after the planning and the scheduling processes. For instance, steps b-g further involve data associated with at least one of the one or more additional processes, such as the distribution process described herein. For example, the sets of parameters include any combination of the first, second, third, and additional sets of parameters, the inputs include any combination of the first, second, third, and additional inputs, the outputs include any combination of the first, second, third, and additionaloutputs, the KPIs include any combination of the first, second, third, and additional KPIs.Further examples include, estimated KPIs including any combination of the first, second, third, and additional estimated KPIs, estimated outputs including any combination of the first, second, third, and additional estimated outputs.

[0117] Furthermore, the processes run in the ACS are controlled by a controller 313 controllingthe set of parameters by which the processes are specified. For instance, although not shown in Fig. 3B, the planning and scheduling processes are controlled by a controller 313 controlling the first and second parallel set of constraint parameter(s) P1,1c, P1,2c and algorithmic parameter(s) P1,1a,, P1,2a and the second set of parameters P2c, P2a. In other instances, where Figs. 3A-3C include one or more additional processes, such as the distribution process described herein, the set of parameters of said one or more additional processes are also controlled by the controller 313.

[0118] In other examples, the ACS may be operated in a particular parallel structure furtherincluding running one or more additional processes, such as one or more distribution processes,in parallel. In the example of the parallel input-output structure, a plurality of one or moreadditional processes is run in parallel, where the number of each of the one or more additionalprocesses is equal in number to the plurality of planning and scheduling processes. In theexample of the parallel multi-input structure, one or more additional multi-input processes areintegrated with the planning and scheduling processes, e.g., run in parallel after the plurality ofplanning processes, such as before and / or after the scheduling process.

[0119] In embodiments according to the invention, the method includes receiving an observedresult of an execution of the generated (and preferably transmitted) execution plan. The observed result may at least one of an observed performance indicator KPIOb and an observedoutput OOb. Said observed result may be received from the AES as described herein or fromany other execution system which can provide said observed result for a given execution plan(i.e., input demand and sets of parameters).

[0120] The one or more datasets may be enriched with predicted data points generated byapplying a machine learning model to a generated input demand, wherein the machine learningmodel is learned on combined properties of the planning and the scheduling processes. Theproperties may include at least one of: definition of the parameters involved (i.e., inputs,outputs, the set of parameters, KPIs), costs, skills, timings and how said at least one parameterrelates to industrial equipment and resources. In example 1 below, the advantage of themachine learning model for enriching the one or more datasets is shown. The machine learningmodel may be one of a combination of a clustering and regression techniques.

[0121] The one or more datasets may be completed with predicted data points generated byapplying the machine learning model to existing data points, wherein the machine learning extrapolates missing data points. For example, the machine learning model predicts a missing first output of the first solver. It is more interesting that the machine learning model predicts amissing estimated result and / or a missing observed result. By including a clustering techniquein the machine learning model, knowledge can be incorporated over the topology of theproblem space, and more precisely, over the different critical regions that may exist in thesolution space. The machine learning model can be used to quickly assess the performance ofan optimal system in face of unseen demand instances. In one example, this could be used tohelp in completing the one or more datasets, filling in gaps where KPIs for a given tuple (input,output, KPI) have not been assessed.

[0122] In step c (i.e., determining 104), it is highly preferred that the calculation of distancesbetween the performance indicators in the one or more datasets and the desired combination ofthe performance indicators includes calculating a weighted distance between the performanceindicators in the one or more datasets and the desired combination of the performanceindicators. Performance indicators that have been observed are provided with a higher weight than performance indicators that have been estimated which are provided with a higher weight than the performance indicators that have been predicted. This ensures that more weight is provided to more accurate KPIs, and thus, ensure that the combined set of values for the setsof parameters is more effectively determined and would result in a an execution plan that wouldmore accurately resemble the desired combination of KPIs.

[0123] The data points in the one or more datasets may be included in vectors, wherein eachvector includes data points each including data of the performance indicators and data of thetherewith associated inputs, outputs, and sets of parameters. Thus, each vector may include the historical data points as described herein. Furthermore, each vector may include predicted data points as described herein.

[0124] Fig. 2 illustrates an assembly configuration system (ACS) comprised in areconfiguration assembly system (RAS). The ACS 210 comprises means (e.g., a processor 211) for carrying out any of the embodiments described herein. As shown in Fig. 2 the ACS 210further comprises one or more datasets 212 communicatively coupled to the processor 211.The processor 211 may perform any one or more of the following: retrieve datapoints from theone or more datasets 212, store datapoints in the one or more datasets 212 as described herein,enrich the one or more datasets 212 as described herein, etc. Thus, it will be understood thatthe means (e.g., the processor 211) performs the method of operating the ACS as described herein.

[0125] The processor 211 or the controller 313 may be connected to the solvers of the processesand / or to the AES through a communication bus and protocol (e.g. AutomationML). The processor 211 or the controller 313 monitors at least one of: the estimated KPI(s) and the observed KPI during assembly operation or execution. Additionally or alternatively, the processor 211 or the controller 313 monitors at least one of: the estimated output(s) and the observed output during assembly operation or execution. The processor 211 or the controller 313 may further monitor a service specific KPI, such as solver computational time. The processor 211 or the controller 313 may start / stop the solvers and control the operation of AES. The processor 211 or the controller 313 may assign values to the set of parameters of the solvers by sending said values to the solvers.

[0126] In embodiments according to the present invention, the controller 313 used forcontrolling the set of parameters as described herein may be the processor 211 included in the ACS 210. In preferred embodiments, the controller 313 is included in the ACS 210 and is communicatively coupled to the processer 211, wherein the processor 211 transmits instructions to the controller, which when executed by the controller, cause the controller to control the set of parameters as described herein.

[0127] The assembly execution system (AES) comprised in the RAS 200 is configured toreceive the generated execution plan from the means (i.e., processor 211) and is configured to execute said generated execution plan. The ACS 210 may receive from the AES 220 anobserved result (e.g., an observed performance indicator KPIOb and / or an observed output OOb)after executing the generated execution plan. As shown in Fig. 2, the observed result is directlystored in the one or more datasets 212 by the AES 220. Alternatively, the processor 211 mayreceive said observed result and store it in the one or more datasets 212. Furthermore, theprocessor 211 may analyze the observed result to tune the parameters of the processes.

[0128] The data points in the one or more datasets may comprise historical data pointsincluding observed data points and optionally estimated data points. The observed data points may be observed by the AES of the RAS or by other execution systems which can provide saidobserved result for a given execution plan (i.e., input demand and sets of parameters). Theestimated data points include data of the estimated KPIs and / or estimated outputs based on an input demand and the assigned set of parameters as described herein.Examples1. Example 1

[0129] In this first example, the case of aggregate planning for multi-product assembly systemswith reconfigurable cells is considered, particularly that of Uzunosmanoğlu, Mehmet, et al(2021) (DOI: 10.1007 / 978-3-030-85902-2_56). The aim is to compare state-of the art MachineLearning (ML) models to the ML model of the present inventors for predicting a high-levelsummary of the optimal solutions obtained when solving an Integer Quadratic Problem (IQP)to optimality. Here, a single time-period operation is assumed. Finally, a high-level summaryof an optimal solution to the IQP problem for a given instance of demand is defined as containing (i) the minimal set of value-adding resources to be allocated to the system and (ii) the minimum total cost of ownership associated to the optimal system (that with an optimal cell configuration, system configuration and product routing).

[0130] For the prediction of the high-level summary of an optimal solution a ML metamodelas described herein (hereinafter called the ML sampler) is defined to include clustering andregression techniques to improve the performance of off-the-shelf ML models. A total of 7000samples are collected, of which, 80% (5600 samples) are used for training and testing and ofwhich 20% (1400 samples) are used for validation. Additionally, the performance of the MLsampler in predicting a correct high-level summary of the optimal solution to each problem instance is compared with that of commonly used off-the-shelf models, namely: (i) Linear Regression (LR), (ii) K-nearest-neighbors regression (KNN), (iii) Ensemble of extremely randomized decision trees (ET). All the off-the-shelf models have been optimized using a grid search with 5-fold cross-validation (on the training set). Note that a “Baseline” regressor isincluded which corresponds to a dummy regressor that always predicts the mean of the trainingset.

[0131] The first component of the high-level summary is the minimal set of value-addingresources to be allocated to the system. Note that the cost of acquiring a component is a bigpart of the total cost of ownership. At the same time, it greatly constrains the maximum capacityof the system. Which means that an underestimation of the number (and / or types of resources) could lead to a solution that is infeasible, i.e., the set of resources is not able to assemble 100%of the product units in 1-time period. The solutions obtained can be seen in the table below:Method Optimal solutions Infeasible solutions CPU TimeML sampler 90,50% 2,07% 3,1 minutes*Aggregate solver 100% 0% 11 hoursBaseline regressor 40,29% 31,29% 0,001 secs*LR 54,86% 19,14% 0,009 secs*KNN 76,29% 8,21% 0,13 secs*ET 78,36% 7,29% 9,84 secs**Including training

[0132] The best performant off-the-shelf method is ET which achieves a 78% accuracy in thepredictions of the resource set with 7% of the predicted sets corresponding to infeasible solutions. In contrast, the ML sampler can predict the minimal set of value-adding resources to be allocated to the system more than 90% of the times, and only in about 2% of the times the prediction leads to resource set incapable of assembling 100% of the products in the specified time (infeasible solutions). This corresponds to an increase in performance (wrt to the ET method) of 15.5% in predicting correctly the optimal set of solutions and 71.6% less infeasible solutions with the ML Sampler method. The difference in time seems big, but it must be reminded that thanks to parallelism, these problems scale linearly with size.

[0133] For the second component of the high-level summary of the solution, the KPI thatrepresents the minimum total cost of ownership associated to the optimal system. Note that for the evaluation of the different methods we use the following metrics, commonly used to assess the predictive performance of a predictive method: Mean absolute error (MAE), Mean absolute percentual error (MAPE), Maximum absolute percentual error (Max PE), and Root-Mean- square error (RMSE). The solutions obtained in the experiments can be seen in the table below: Method MAE MAPE Max PE RMSE CPU TimeML sampler 48,21 0,40% 7,19% 88,22 38,26 secs*Aggregate solver - - - - 11 hours*Baseline regressor 2316,08 22,10% 367,97% 2972,45 0,001 secs*LR 3250,76 28,87% 559,88% 269,75 0,009 secs*KNN 3159,74 28,21% 545,74% 576,74 0,12 secs*ET 294,64 2,34% 26,75% 399,10 6,57 secs**Including training

[0134] Once again, ET is the best performant off-the-shelf method. However, When comparedwith the ET, the ML sampler reduces error as measured by the MAE by 84%, the MAPE by83%, the Max PE by 73%, and the RMSE by 78%. All in all, the experiment shows that theML sampler has a significantly better performance than other off-the-shelf ML models, andthat it can be used to predict high level summary solutions of optimal solutions, more precisely to predict the optimal objective value of a parameterized optimization problem, with a good performance.2. Example 2

[0135] In this second example, the present method for operating the ACS is compared to thestandard use of the solver in a hierarchical control scenario, i.e., the rolling horizon handover.

[0136] For the use case at hand, the present method for operating an ACS makes use ofprecalculated aggregated samples in such a way that we can derive the best set of parameters P1c from the data to solver S1 in face of a new realization of demand. In this experiment, changes to P1c will be aimed at those constraint parameters that influence the resource allocationvariables in the model. Note that the parameters (P1a, P2a) will not be considered / modified sincethe algorithmic parameters will only affect the time that it takes for the solvers to reach to an optimal solution. Parameter set P2c is also left untouched (default values).

[0137] The inventors have explored the case in which the demand (daily orders) distributionof the products is independent and identically distributed (i.i.d) and highly volatile. Herewith,a study is provided on the demand during the course of approximately 2 years (720 days),where the state of practice approach (i.e., the rolling horizon planning) and the present method of operating an ACS are compared in production planning, i.e., process planning (equipment selection, system configuration, routing), and production scheduling.

[0138] The results showed that:- The present method delivers a better performance calculated in terms of OTD rate thana rolling horizon planning strategy. Average OTD (both measured weekly or monthly) increases by a margin of 6%. In some months an increase in OTD / month of as much as30% can be observed (or as much as 57% for OTD / week). This is achieved without significant changes to the realized TCO in the same period of time.- The present method incorporates a higher level of nervousness (cf. number of resourcepool changeovers) into the system, meaning that the system reacts more frequently tochanges in demand: The present method results in a total of 33 resource pool changeovers vs 13 when using rolling horizon planning. Nervousness refers to a lackof planning stability. -Total CPU time is drastically reduced with the present method. Rolling horizonplanning takes as much as 26 times more time in total (183 min vs 7 min).3. Example 3In this third example, the present method based on the use of the proposed mapping process is compared to conventional (black box) methods. Differences include, ease of model update due to modularity of the method and reduction in problem space (DB size) with respect to black box methods. In particular, this example stresses the importance of presence of O1 and I2 in the data structure of the recommender database. This specific property allows the present method to generate theplanning decision space as (I1,O1) and the scheduling decision space (I2,O2). The planningand scheduling process is as depicted in Figure 4.Planning and scheduling process are each carried out by solving an optimization problem via a given solverand ^^, respectively. Each solver (i) has model / control parameters (^^^) and algorithm parameters (^^^), inputs (^^), outputs (^^) and KPIs (^^^^). These parameters influence the output of each solver for a given input. Note that the planning-scheduling process is sequential and hierarchical and as such, the output of solverinfluences the input of solver ^^. The output of the full planning-scheduling process corresponds to the combined output ofboth solvers, i.e., ^^ ∪ ^^.Now considering the case in which the solverdeals with a resource allocation problemsimilar to that in (Bretthauer & Shetty, 1995), whilst ^^deals with a flexible job shop scheduling problem similar to that in (Shen et al., 2018). The resource allocation problem in ^^is frequently found in production planning and in this case deals with identifying the output which corresponds to the set of machines that need to be allocated to produce a given demand or order list (^^). The set of resourcesdefined inare then used as input, together with a given demand ^^, in the flexible job shop scheduling problem in ^^where output ^^defines how jobs are to be allocated to machines and the timings for the execution of these jobs.In this example, an assembly system for a product mix of 5 products with a demand for eachproduct in the range of 16 to 32 units is provided. Assembly of each product is defined by afixed set of assembly operations. There are in total 5 machine types with each machine suitable for at least 1 assembly operation type for at least 1 product type. The production times and costs for a given machine for a given task of a given product are fixed. At least 2 machines of a different kind each are needed to produce any given order, and at most 8 machines can beused at any given time. Furthermore, solver ^^includes as a model parameter (^^^) the “slack” parameter ^^which multiplies the total capacity of a given machine type ^^, so that the capacityof machine ^ is represented by ^^ ∗ ^^ , we consider |^^^| = 32. Solver ^^ includes as a modelparameter (^^^) the parameterthat defines the setup time incurred when tasks of jobs ^ and^ are process sequentially in a given machine of type ^, with ^^^^ = 0 and ^^^^ = ^^^^ , weconsider |^^^| = 24.The ^^^^represents the total production cost and corresponds to the objective value of the resource allocation problem as determined by solver ^^. ^^^^will correspond to the expected makespan and is obtained via discrete event simulation (DES) using historical setup time data. Both ^^and ^^are NP-Hard optimization problems and will be solved using Gurobi (GurobiOptimization LLC, 2023) solver. To solve these problems, Gurobi exposes algorithmparameters that modify its performance. These parameters control the way Gurobi operates inorder to find a solution to the problem. We follow the work of (Nau et al., 2022) whichenumerates the 3 crucial algorithm parameters (^^^ , ^^^) for Gurobi, namely the parameters“Cuts”, “MIPFocus” and “Presolve”, we consider then |^^^| = |^^^| = 36.It is clear that both model and algorithm parameters influence the performance of the combined solver output, e.g., different values of the “slack” parameter ^^of solvercan lead to different types and number of machines eventually allocated; since solver ^^is deterministic but ^^^^is defined via an stochastic process (DES), different values of the setup time parameter estimate ^^^^used in solver ^^will lead to different schedules, each with different performance ^^^^; finally, for a given maximum CPU time allocated for the execution of each solver, different configurations of the algorithm (Gurobi) parameters may lead to different solutions (and respective KPIs) for all timed out instances. Consequently, for a given demand or order list ^^, optimal control (understand as a planning-scheduling process execution with optimal performance) will correspond to the selection of both model and algorithm parameters(^^^ , ^^^ , ^^^ , ^^^) to be used in the execution of planning and scheduling solvers (^^, ^^) thatoptimize the performance (^^^^, ^^^^) of the combined solver output (^^, ^^), or atransformation of the performance via a given known function ^(^^^^, ^^^^).The surrogate models derived from the mapping process can then be used to cheaply identifythe joint performance (^^^^, ^^^^) of the planning-scheduling process represented by solvers(^^, ^^) for a given parameter set (^^^ , ^^^ , ^^^ , ^^^) in the face of a given raw input demand(order list) ^^by means of the following steps: 1. A new incoming order (list) is received as raw input ^^.2. Raw input ^ in step (1) is ma^ ^pped to the most similar input family ^^. In the current example, this is done by classifying the demand of each of the 5 products in the product mix as either low, medium, or high demand. As such, ^^ ^^^ = 243.3. Surrogate model i^s used to evaluate parameters ^^^ , ^^^ for the order family ^^identified in step (2). The outcome is a related output family ^^^and associated KPIs (^^^^). Note that an output family ^^^corresponds to a representation of a solution to solver ^^and not the explicit solution ^^. In the current example, ^^^corresponds to a set of features that describe output ^^, the resulting features describe the range of the total number of machines allocated (the ranges considered are [2,4), [4,6),[6,8]), and a boolean variable for each machine type that is allocated, i.e., a 5-dimensional vector that denotes whether a given machine type (out of the 5 machine types available) is allocated (1) or not (0). As such,= 72.4. Mapping function ^ is used to map all the unique output families ^^^identified in step (3) to corresponding input families ^^^. The input families ^^^are each a unique representation of the input ^^ =to solver ^^ that defines an instance of a flexiblejob shop scheduling problem. In the current example, ^^^ = ^^ ^∪ ^^^is a representation of the associated problem instance ^^^ ^. As such, |^^ | = ^^^ × ^^^^. 5. Surrogate model ^^^is used to evaluate ^^^ , ^^^for the input family ^^^identified in step (4). The outcome is a related output family ^^^with associated KPIs (^^^^). Note that in the current example is not necessary to define the output family ^^^, in fact, surrogate model ^^^can deal only with estimating the KPIs (^^^^) associated to a given tuple (^^^ , ^^^ , ^^^). The definition of an associated output family ^^^is only relevant if we were to extend the model to incorporate a third optimization process ^^that is influenced by the second optimization process ^^. By following the method steps described above one can cheaply explore a large variety ofmodel and algorithm parameters (^^^ , ^^^ , ^^^ , ^^^) in order to find the parameter configurationthat yields best joint performance (^^^^, ^^^^). This process can be done manually orautomatically (using, e.g., local search method such as gradient descent, or global search methods such as genetic algorithms). Note that the key difference between the model here proposed and other black box approaches is the consideration of the intermediate outputs (^^) and their mapping to the intermediate inputs (^^). Whilst a black box approach (see below) will normally only focus on the first input (^^). This key difference leads to two main advantages:-It divides the attention from joint global performance to the performance of theindividual solvers. This means that the surrogate models are trained independently from one another, as such, updates or changes to one solver will not require of updates of changes to any other solver that is part of the sequential process. -It can lead to significant reductions in the size of the problem space, even if we considerthat similar input clustering is done in the in the black box approach. The black box approach will deal with a problem space defined by ^ ^ ^× ^^^ × ^^^ × ^^^ × ^^^ whilstthe proposed method deals with 2 separate problem spaces ^ ^ ^× ^^^ × ^^^ and^^^ × ^^^ × ^^^ for the surrogate models ^^^and ^^^, respectively. If we consider the problem space of the proposed method as the sum of the problem spaces for the independent surrogate models, the problem space of the black box method will consist of= 241^864.704,00 problem instances. Theproblem space of the proposed method will consist of |^ ^ ^| × |^^^| × |^^^| +|^^^ | × |^^^| × |^^^| = 15^396.480,00 problem instances. Meaning that the proposedmethod leads to a 93.6% reduction in the size of the problem space with respect to theblack box approach.4. Example 4In this fourth example, in addition to example 2, the current example provides detail on theapplication of the method in a similar use case and compares it with the situation in which no control method is applied.Considering the case of example 3 in which two solversand ^^ are used sequentially forplanning and scheduling in a reconfigurable assembly system (RAS), see Figure 5. Solver ^^with the definitionnamely the (i) types and number of resources to allocate, (ii) the physical distribution of these resources in the shopfloor (layout) and (iii) the explicit allocation of assembly tasks to resources in the shopfloor (routing). Solver ^^deals with ^^, namely the scheduling of the assembly tasks based on the routing defined by solver ^^. The input (^^) to solver ^^corresponds to the demand of products that need to be assembled, whilst the input (^^) to solver ^^corresponds to the routing defined by solver ^^. The key performanceindicators (KPIs) for solversand ^^ are ^(^^) = ^^^^ and ^(^^) = ^^^^, respectively.Note that ^^^^denote cost-based KPIs whilst ^^^^are time-based KPIs. Solverhas model / control parameters (^^^) that describe limits on allocated capacity (slack) of resources, and bounds on the number of resources to allocate. Solver ^^has model / control parameters (^^^) that describe sequence dependent changeover times. Furthermore, solveris solvedexactly using Gurobi solver (Gurobi Optimization LLC, 2023) with algorithm parameters (^^^) as defined by Gurobi. Solver ^^is a greedy iterative algorithm with algorithm parameters (^^^) that define different strategies to update the solutions after each iteration as well as initializationstrategies. These parameters influence the output of each solver for a given input. Theseparameters are usually tuned by the developers of the solvers so that the “default” parameter values achieve the best solver performance (on average) for a given input.Considering the distribution of orders for a product mix of 2 products as shown in Figure 6:mix p1 – p2 = 96% – 4% & coefficient of variation p1 – p2 = 5% – 76%.

[0139] If the parameters of the solvers are left at their default values, the planning-schedulingprocess will yield a solution that is of minimal cost (^^^^) but not necessarily of minimal time (^^^^) since the output of ^^is conditioned by the output of ^^. For this reason, optimal solver control will generally aim to find a balance between the conflicting views of each individual solver. Proper selection of solver parameters can lead to optimal solver control in this case.From mapping process, we derive a surrogate system for the planning and scheduling processthat can be used to cheaply identify optimal solver parameters. This procedure can be applied at any given point in time to, for example, find a planning-scheduling solution that prioritizes On-time delivery (OTD) maximization (^^^^) over Total Cost of Ownership (TCO) minimization (^^^^). As such, the method could be used as follows: 1. A new incoming order (list) is received as input ^^^. 2. Order ^^in step (1) is mapped to its closest input fami ^ ^ ly ^^which represents a family of closely related orders. For the case at hand we consider a total of ^ = 10 inputfamilies. 3. Surrogate modelis used to explore the parameter space ^^^ , ^^^for the order family ^^identified in step (2). This to identify the set of all unique output families (^^^) and associated KPIs (^^^ ^ ^) that can be derived for the order family ^^. Note that an output family ^^^can be nothing more than a unique representation of a solution to solverand not the explicit solution ^^. In this case, e.g., it corresponds to a set of features that describe a routing solution ^^, such features include the utilization of the bottleneck resource, the length of the longest path, the bottleneck resource identifier, average path length, among others. For the case at hand we consider a total of ^ = 62 outputfamilies. 4. Mapping function ^ is used to map all the unique output families ^^^identified in step (3) to corresponding input families ^^^. The input families ^^^are each a unique representation of the input ^^(which corresponds to an explicit routing) to solver ^^. In this case, e.g., ^^^corresponds to a subset of the features that describe a unique output family ^^and the feature ^ ^ s that describe the pertinent input family ^^. Note that in this specific case ^ is bijective and as such there are ^ = 62 input families for the surrogatesolver ^^^. 5. Surrogate model ^^^is used to explore the parameter space ^^^ , ^^^for all the input families ^^^identified in step (4) with the goal of identifying all combinations of that yield the best possible KPIs (^^^^), i.e., maximal OTD. 6. We select from the subset of output families ^^^identified in step (3) the ones associated with the input families ^^^which maximize ^^^^identified in step (5). From this subset of output families, we select the output family ^^^(and associated order family ^ ^ ^ and solver parameters ^^^ , ^^^) that yields the best possible KPIs (^^^^), i.e., minimal TCO.7. The method returns the solver parameters ^^^ , ^^^identified in step (6) and the solver parameters ^^ , ^^associated to the input f^^ ^^amily ^^that maps the output family ^^(and associated order family ^ ^^^and solver parameters ^^ , ^^^) identified in step (6) so that ^^^ , ^^^ , ^^^is a combination identified in step (5). Following the method described above, we find a planning-scheduling solution that prioritizes On-time delivery (OTD) maximization (^^^^) over Total Cost of Ownership (TCO) minimization (^^^^) for the given product mix: Solver parameters Default valuesOptimal controlOTD (^^^^) 0,61 +47%TCO (^^^^) 3739,27 +39%Makespan (^^^^) 12032,28 -33%# late units (^^^^)8,89 -73%Optimal control via proper parameter selection can lead in this example to a planning- scheduling solution that increases OTD by 47% compared to the solution obtained with the default parameter values. In this case, optimal control defines a preference of OTD over cost,however, it is not unclustered for these preferences to change over time to, e.g., adapt toexisting market conditions, or due to changes in tactical or strategic objectives. As a result, solver parameter selection should be able to map the different preferences from users. This is easily achieved in this case since the surrogate modelsand ^^can be used independently,and a function ^(^^^^, ^^^^) that maps any optimal control preference can be evaluated atruntime.5. Example 5In this fifth example, the application of the proposed method as a part of an operational BImodule is explained. Advantages of the integration of the method in such modules include: decision-making support, and responsiveness.The present method has been incorporated in a custom Operational BI application. Theapplication provides, among other things, with a Dashboard that showcases the assembly system’s performance from the point of view of the involved solvers (Planning & Scheduling solvers), simulation / digital-twin validation applications (or extended to see online performance): In the application above the Key Performance Indicators (KPIs) presented are calculated directly from the solver’s solution and / or reflect the objective value of the solution found by the solvers. It can be extended, however, to involve custom methods for KPI determination on the solution found. The Operational BI incorporates the proposed method in 2 ways: 1. It can be used to control the solvers automatically and dynamically via their modeland / or algorithm parameters. This can be done by selecting the “Data-based operation mode” in the settings menu.2. It can be used to automatically and dynamically present suggested solver configurationswith different performance over the KPI set. As such, the user can manually select the configuration that best fits the decision-making at a given point in time. As such, the proposed method can provide users with increased responsiveness achieved by the agile solver control provided by the “data-based operation mode” since optimal solver parameters are identified from effectively instantaneous predictions leveraged on the proposed method. The fast response of the method can also be applied to quick scenario analysis, this can be manually performed but is also automatically carried on via the suggestion / recommendation alerts leveraged on the proposed method. Which in combination with the dashboard view facilitates informed decision-making. Finally, the BI module can be extended to accommodate 2 or more assembly lines providingholistic shop floor control as shown in Figure 3D. In cases where the same solvers are used andsimilar assembly contexts are considered for a set of parallel lines. An extension of the BI module featuring the proposed method can be achieved by using the underlying DB across the parallel lines interchangeably:In such settings, the dataset used for the control of the solvers of one of such lines can be used interchangeably for the control of the same solvers in the other (parallel) line that operates under a similar assembly context. The simplification can be expanded to account for 2 or more parallel lines and / or solver steps.Original Claims (as filed in the priority application): these original claims are being retained for reference purposes 1. A computer-implemented method for operating an assembly configuration system,ACS, wherein operating the ACS includes running integrated processes of a planning process and a scheduling process, wherein the planning process determines the components and resources, operations and their routing that are needed to manufacture a workpieces into a parts or a products and wherein the scheduling process determines the order in which tasks are executed and the detailed routings over the ACS, wherein the planning process is optimized by a first solver (S1) which determines a first output (O1) from a first input (I1) and which is evaluated by a first performance indicator (KPI S1), wherein the scheduling process is optimized by a second solver (S2) which determines a second output (O2) from a second input (I2) and which is evaluated by a second performance indicator (KPI S2), wherein the first solver (S1) associated with a first set of parameters (P1c, P1a) and the second solver (S2) associated with a second set of parameters (P2c, P2a), wherein the second input (I2) is based on the first output (O1), wherein the planning and scheduling processes are controlled by the first and second set of parameters (P1c, P1a, P2c, P2a), respectively, the method comprising: a. receiving (101) one or more input demands as the first input (I1);b. receiving (102) one or more datasets including data of the performanceindicators (KPI S1, KPI S2) and data of the therewith associated inputs (I1, I2), outputs (O1, O2), and sets of parameters (P1c, P1a, P2c, P2a); c. determining (104) from the one or more datasets a combined set of values forthe parameters to be (re)configured (P1c, P1a, P2c, P2a) based on the first input (I1) and on a desired combination of the performance indicators; d. configuring (106) the sets of parameters (P1c, P1a, P2c, P2a) by assigning thecombined set of values thereto; and e. generating (108) one or more execution plans based on the first input (I1) andthe configured sets of parameters (P1c, P1a, P2c, P2a), said execution plan comprises the set of resources needed to produce the input demand as the firstoutput (O1) and a schedule for that set of resources that schedules the jobs over the resources as the second output (O2). 2. The method according to claim 1, wherein the one or more datasets comprise:- a first dataset including data points each including data of the performanceindicator (KPI S1) and data of the therewith associated input (I1), output (O1), and a set of parameters (P1c, P1a); and -a second dataset including data points each including data of the performanceindicator (KPI S2) and data of the therewith associated input (I2), output (O2), and a set of parameters (P2c, P2a), wherein the step of determining a combined set of values for the parameters to be (re)configured (P1c, P1a, P2c, P2a) from the datasets comprises: i. determining a set of values for the parameters to be (re)configured of the firstsolver (P1c, P1a) and the first output (O1) from the first dataset based on the first input (I1) and on a desired performance indicator; ii. mapping the first output (O1) from the first dataset to a second input (I2) fromthe second dataset; and iii. determining a second set of values for the parameters to be (re)configured ofthe second solver (P2c, P2a) and the second output (O2) from the second datasetbased on the second input (I2) and on a desired performance indicator.The method according to claim 2,wherein the data points in the more datasets are grouped in groups of data pointsassociated with the similar in- and output values,wherein each groups of data points is represented by a clustered data point including aninput index representing the distribution of input values in the group, an output indexrepresenting the distribution of output values in the group, a KPI index representing adistribution of values of KPIs associated with the distribution of inputs and outputs anda set of values for parameters (P1c, P1a) associated with the distribution of inputs and outputs and KPI values, wherein the combined set of values for the set of parameters (P1c, P1a, P2c, P2a) are determined based on the clustered data points.The method according to claim 3, wherein the mapping step comprises mapping thefirst output (O1) of a clustered data point from the first dataset to a second input (I2) ofa clustered data point from the second dataset.The method according to claim 4,wherein the mapping step comprises mapping the output index to a second input indexof a clustered data point from the second dataset.The method according to any one of the preceding claims, wherein the method furthercomprises receiving a set of initial configured parameters,wherein determining (104) the set of values for parameters to be (re)configured (P1c,P1a, P2c, P2a) is also based on the set initial configured parameters.The method according to any one of the preceding claims,wherein the data points in the one or more datasets comprise historical data points including observed data points and estimated data points and further includes predicted data points.The method according to claim 7,wherein the predicted data points are generated by applying a machine learning model to a generated input demand, wherein the machine learning model is learned on combined properties of the planning and the scheduling processes, wherein more preferably the generated input demand is a classification of the input demand.The method according to any one of the preceding claims, wherein operating the ACSincludes running a plurality of planning processes in parallel, wherein the plurality of planning processes share a resource pool, wherein operating the ACS preferably includes running a plurality of scheduling processes in parallel, wherein the plurality of scheduling processes is equal in number to the plurality of planning processes, wherein the input of each of the plurality of scheduling processes is the output of the respective of the plurality of planning processes.The method according to any one of the preceding claims, wherein operating the ACSincludes running a distribution process optimized by a third solver (S3) which determines a third output (O3) from a third input (I3) and which is evaluated by a thirdperformance indicator (KPI S3), wherein the third solver (S3) associated with a third set of parameters (P3c, P3a), wherein the distribution process is integrated with the planning process and the scheduling process such that the third input (I3) is the first output (O1) and the second input (I2) is the third output (O3), wherein the distribution process is controlled by the third set of parameters (P3c, P3a).The method according to any one of the preceding claims, further comprising:f. simulating (110) the generated execution plan to generate an estimatedperformance indicator, preferably a combination of estimated performance indicators; g. calculating (112) a distance between the estimated performance indicator and aperformance indicator from the desired combination of the performance indicators, preferably between the combination of estimated performance indicators and the desired combination of the performance indicators; and h. when the calculated distance exceeds (114) a predetermined threshold,repeating steps c to h until it is determined that the calculated distance does not exceed the predetermined threshold.The method according to any one of the preceding claims, wherein step e) comprisesgenerating at least two different execution plans, and prompting a user to select a desired execution plan out of the at least two different execution plans, preferably based on a desired combination of the performance indicators.An assembly configuration system, ACS, comprising:- a means (211) for carrying out the method of any one of claims 1-12; and- preferably one or more datasets (212) communicatively coupled to saidmeans.A reconfiguration assembly system, RAS, comprising:- an assembly configuration system, ACS, (210) according to claim 13; and- an assembly execution system, AES, (220) configured to execute anexecution plan generated by the ACS (210) according to any one of claims 1-12.15. A computer-readable medium comprising instructions which, when executed by acomputer, cause the computer to carry out the method of any one of claims 1-12.

Claims

1. Claims 16. A computer-implemented method for determining, from one or more datasetsincluding data of the performance indicators KPI S1, KPI S2and data of the therewith associated inputs I1, I2, outputs O1, O2, and sets of parameters P1c, P1a, P2c, P2a, a combined set of values for a set of parameters wherein the parameters influence the output of a solver for a given input and dynamically control performance of planning and scheduling processes of an assembly configuration system, ACS, by generating one or more execution plans under the values for the sets of parameters, the computer-implemented method comprising the steps of: -determining a combined set of values for parameters to be reconfigured from the oneor more datasets by (i) determining a set of values for the parameters to be reconfigured for a firstsolver and a first output from a first dataset that is based on the given input,(ii) determining a partial set of configured parameters, -and, for a desired performance indicator, mapping the first output from the firstdataset to a second input of a second dataset to determine a second set of values forthe parameters to be reconfigured for a second solver and mapping a second outputfrom the second dataset based on (i) the second input, (ii) the partial set of configuredparameters, and (iii) a desired performance indicator.

17. The method of claim 16 further comprising the steps of:grouping data points in the one or more datasets wherein each group of data points is represented by a clustered data point that represents a distribution of: input values, output values, andKPI’s that reflect the relation between input and output for a given parameter set,and wherein the clustered data point is stored as a clustered value for the input, a clustered value for the output, a KPI set and a parameter set.

18. The method of claim 17 further wherein the mapping step comprises:mapping the first output O1 of a clustered data point from the first dataset to a second input I2 of a clustered data point from the second dataset.

19. The method according to claim 16 or 17 wherein the data set is enriched withpredicted data points generated by applying a machine learning model to a generated input demand.

20. The method according to claim 17 or 18 wherein a machine learning module isapplied to the generated input demands allows enriching of a dataset byincluding more data points.

21. The method of any claim 16 – 18 further comprising:a dataset is enriched with predicted data points generated by applying a machinelearning model to a generated input demand, and wherein the machine learning modelis taught on combined properties of a planning process and a scheduling process.

22. The method of claim 16 wherein the step of determining a combined set ofvalues for the parameters to be reconfigured further comprises: selecting the performance indicators in the one or more datasets closest to the desired combination of the performance indicators.

23. A method of running integrated processes of a planning process and ascheduling process of an assembly configuration system, ACS, the methodcarried out by a computer and comprising the following steps: controlling the planning process by a first solver S1 by reconfiguring a first set ofparameters by determining, for a first problem instance I1, a closest clustered input I1jand retrieving an associated first performance indicator and values for a firstparameter set I1j, KPI1, P1c, P1a;retrieving, for the planning process by a first solver S1 for a first problem instance I1and a first parameter set P1a,P1c, a first clustered output O1k by determining for afirst problem instance I1 a closest clustered input I1j and retrieving the clusteredoutput O1k associated to the clustered input I1j and the first parameter set P1a,P1c;controlling the scheduling process by a second solver S2 by reconfiguring a secondset of parameters by determining for a first problem instance I1 and the first parameterset P1a,P1c and their associated first clustered output O1k, a second closest clusteredinput I2m and retrieving an associated second performance indicator and values forthe second parameter set I2m, KPI2, P2c, P2a;wherein the first clustered output O1k corresponds to a set of machines that need to beallocated to produce a given demand defined by the first problem instance I1; and thefirst clustered output O1k is then used as input, together with the first probleminstance I1 to retrieve the second parameter set of the second solver S2; andthe running integrated processes of the planning process and the scheduling process iscontrolled by a combined parameter set P1c, P1a, P2c, P2a that controls the output ofthe first solver S1 and the second solver S2 to produce output of a given demand.

24. The computer implemented method of claim 23 further comprising thefollowing steps carried out by a computer of: generating a planning decision space I1, O1, KPI1 and a scheduling decisionspace I2, O2, KPI2.

25. The computer implemented method of claim 23 or 24 wherein the first solverS1 is a resource allocation problem and the second solver S2 is a flexible jobshop scheduling problem, the method further comprising the following steps carried out by a computer: outputting by the first solver S1 and the second solver S2 as an output thatcorresponds to a set of machines that need to be allocated to produce a given demandor order list wherein the output defines how jobs are to be allocated to the set ofmachines and timings for execution of the jobs on the set of machines.

26. The method of any claim 23 – 25 wherein the first solver S1 furthercomprises, as a model parameter P1c, a slack parameter which multiplies atotal capacity of a given machine type so that the capacity of the givenmachine is represented by Zk ∗ bk.

27. The method of any claim 23 – 26 further comprising the computer-implemented method steps of:identifying joint performance KPI1, KPI2 of the planning-scheduling process represented by the first and second solvers S1, S2 for a given parameter set P1c, P1a,P2c, P2a given the problem instance or order list I1;receiving a new incoming order list I1 as raw input;mapping raw input I1 to a closest clustered input I1j by classifying the orderlist as either low, medium, or high demand; evaluating, using a first surrogate model S1’, first parameters P1c, P1a for theclustered input I1j to obtain an output family O1k and associated KPIs KPI1; andmapping, using a mapping function H, the output family O1k to correspondinginput families I2m.

28. The method of claim 27 further comprising the following steps implementedon a computer: evaluating, using the second surrogate model S2', the second parameters P2c,P2a for the second input family I2m and relating the outcome to the second outputfamily O2n with associated second KPIs KPI2.

29. The method of any claim 23 – 28 further comprising the following stepsimplemented on a computer:finding a parameter configuration that yields desired / target joint performance KPI1,KPI2 using a search method, the search method comprising either - a gradientdescent, or a global search, or a global search consisting of a genetic algorithm.

30. A data carrier signal carrying the computer program of any preceding claim.

Images