GOB forming system to automatically control production of glass gobs

Abstract

A gob forming system (14) for an individual section machine (10), the gob forming system (14) comprising: a gob feeder apparatus (30) configured to receive a flow of molten glass (13) and to generate a plurality of molten glass streams (45); a shearing mechanism (32) configured to receive the molten glass streams (45) and cut each molten glass streams (45) into at least one glass gob (15), the glass gob (15) of each molten glass stream (45) being configured to be received by a respective individual section (18a; 18b) of the individual section machine (10); a gob sensing apparatus (36) configured to measure gob parameters indicative of the glass gobs (15); and a gob forming control unit (34) configured to acquire the gob parameters from the gob sensing apparatus (36) and control the gob feeder apparatus (30) and the shearing mechanism (32) based on control parameters determined based on the gob parameters, wherein the gob parameters comprise at least the gob mass of the glass gobs (15), and wherein the gob forming control unit (34) is configured to implement a model and an optimiser to determine, starting from the measured gob parameters, the control parameters for controlling the gob feeder apparatus (30) and the shearing mechanism (32), the optimiser being a meta-heuristic optimiser.

Description

[0001] "GOB FORMING SYSTEM TO AUTOMATICALLY CONTROL PRODUCTION OF

[0002] GLASS GOBS"

[0003] Cross-Reference to Related Applications

[0004] This Patent Application claims priority from Italian Patent Application No . 102024000003241 filed on February 15 , 2024 , the entire disclosure of which is incorporated herein by reference .

[0005] Technical Field

[0006] The present invention relates to a gob forming system to automatically control production of glass gobs , in particular used for producing glass containers in an Individual Section ( IS ) machine . Moreover, it relates to an individual section machine comprising the gob forming system, to a method for producing the glass gobs and to a related computer program product .

[0007] Background of the Invention

[0008] As known, glass containers are made in a manufacturing process that executes a succession of operations , namely the batch house , the hot end and the cold end .

[0009] The batch house allows to prepare and mix into batches the raw materials required for glass production ( typically sand, soda ash, limestone , feldspar, cullet and other raw materials ) .

[0010] The hot end melts the batched materials into molten glass and produces segments of molten glass ( referred to as glass " gobs" ) that are then moulded into glass containers . It can also anneal the glass containers to prevent them from being weakened due to stresses caused by uneven cooling .

[0011] The cold end inspects the glass containers to ensure that they are of acceptable quality.

[0012] Typically, the moulding phase of the hot end of the manufacturing process is performed in an Individual Section (IS) machine, which comprises a plurality of identical sections (e.g., between two and twenty) , each of which is adapted to produce one or more containers simultaneously (e.g., up to four) .

[0013] In particular, the hot end operations are performed by a furnace, a gob feeder apparatus, a shearing mechanism, a gob distribution arrangement and the sections of the IS machine .

[0014] In the furnace, the batched materials are melted into molten glass and supplied to the gob feeder apparatus. In the gob feeder apparatus, streams of molten glass flow from a feeder bowl through multiple outlets, towards the sections of the IS machine. Each of the streams of molten glass is cut with the shearing mechanism, located below the feeder bowl, into roughly cylindrical segments of glass (i.e., the gobs) , which fall by gravity and are guided by means of the gob distribution arrangement (e.g., comprising scoops, troughs and deflectors) into their respective blank moulds in the respective section of the IS machine. Since each section of the IS machine can have a plurality of sets of moulds which can operate simultaneously, a correspondent number of cut molten glass streams can be simultaneously supplied to these respective sets of moulds.

[0015] In each set of blank moulds (or first moulds) , a precontainer referred to as a parison is formed, either by using a metal plunger to push the glass gob into the blank mould or by blowing the glass gob out from below into the blank mould . Each parison is then inverted and it is trans ferred to a respective blow mould ( or second mould) , where the parison is blown out into the shape of the finished glass container . The blown parison is then cooled in the blow mould to the point where it is suf ficiently rigid to be gripped and removed from the blow mould .

[0016] For example , a known gob feeder apparatus is disclosed in document US7350379 and a known shearing mechanism is disclosed in document EP0712812A2 .

[0017] The control of the system that produces the falling hot glass gobs ( also called in the following gob forming system) is critical to obtain optimal glass containers . In particular, the control parameters used of the gob feeder apparatus must be carefully chosen in order to obtain the target weight , length and shape of the produced gobs , with a high degree of accuracy .

[0018] Therefore , it is desirable to optimise the operation of the gob feeder apparatus and the shearing mechanism to produce glass gobs of uni form quality and features . In fact , the control of the features of glass gobs is of utmost importance in the glass manufacturing industry as it directly impacts the quality, ef ficiency and cost-effectiveness of the production process .

[0019] Current control techniques for glass gobs often rely on empirical rules , manual adj ustments and simplistic mathematical models , often resulting in suboptimal control and limited precision .

[0020] In particular, traditional control techniques often fall short in achieving optimal control of glass gobs features due to their reliance on simplistic controllers and limited ability to adapt to complex, non-linear relationships . These known techniques are typically based on rule-based algorithms or rudimentary mathematical models ( such as linear mapping and Proportional Integral Derivative , PID) , which struggle to capture the intricate dynamics and dependencies involved in the glass manufacturing process . Consequently, suboptimal control decisions can lead to increased material waste , reduced product quality and decreased operational ef ficiency .

[0021] In particular, the known solutions have the following disadvantages .

[0022] For example , when simplistic machine learning techniques are exploited, a common limitation may concern the dataset choice . In fact , current process data on which the dataset is based often lack variations , especially in the case of manual control . Therefore , obtaining a comprehensive dataset that covers the full range of process conditions and variations can be challenging . Additionally, inaccurate or incomplete data lead to suboptimal model performance and negatively af fect the ef fectiveness of the control system .

[0023] Moreover, even with some level of automation, current systems still heavily rely on human operators for decisionmaking and adj ustments . This human dependency introduces two maj or problems . Firstly, current glass manufacturers rely on an ageing workforce without strong renewal , due to the less- than-ideal environment of the glass plant hot end . Secondly, the human dependency adds a degree of variability and potential error that can af fect the quality and consistency of glass gobs . Operator fatigue , shi ft changes , and individual expertise levels can contribute to deviations in control outcomes , reducing the overall ef ficiency and ef fectiveness of the system .

[0024] Moreover, the glass industry can often experience process variations , such as changes in glass composition, environmental conditions and equipment performance . Existing control schemes struggle to adapt to these variations . Therefore , continuous monitoring and periodic updates to the models may be required to ensure optimal control performance under changing conditions .

[0025] Furthermore , current PID control schemes are typically adapted to a single master section . On the other hand, available model predictive control relies on simplistic linear models . These models tend to only converge locally and have limited precision . Additionally, simplistic linear models trained on historical data, as described above , might not generalise well due to unseen scenarios or novel process conditions .

[0026] Consequently, the current automation ef forts have yielded limited results , resulting in noticeable variations and inconsistencies in the weight , length, and shape of glass gobs .

[0027] The limitations of the current control schemes arise from the lack of generalisable data-driven insights and global optimisation techniques . The absence of comprehensive data-driven approaches restricts the ability to achieve precise and consistent control over glass gob properties .

[0028] For example , a known solution about an automatic control that tries to optimise the operation of the gob forming system is disclosed in document EP3611137 . This known solution implements a Model Predictive Control (MPC ) system that combines a simplistic model with an optimiser to control various parameters , such as the movement of plungers of the gob forming system, the timing of shear blades of the shear mechanism and the vertical adj ustment of a cylindrical feeder tube of the gob forming system . Even though this approach of fers a broad application scope , it may not be accurate enough i f suf ficient di f ferentiation cannot be ensured .

[0029] In particular, this solution employs a l inear representation that directly associates variations in the targeted variables (weight and length) with variations in the controlling variables through a matrix of regression ( or sensitivity) coef ficients .

[0030] As evident , this model is based on heavy assumptions such as that a linear mapping exists between the physical parameters of the gobs and the controlling parameters .

[0031] Moreover, this solution is based on the minimisation of the error between the target values and the measured values by means of the solution of an optimisation problem exploiting a cost function . The cost function is represented either as a LI norm or a L2 norm between the desired output values and the sensitivity coef ficient matrix multiplied by the variations of the input vectors . Therefore , this is based on the assumption that an analytical solution can be obtained i f a linear mapping is found and can be solved using standard solvers .

[0032] However, the formulation of the optimisation problem in this known solution is quite general and the speci fic details and constraints of the glass gob control problem are not explicitly considered . These detai ls may strongly l imit the ef fectiveness and applicability of this solution . Furthermore , the reliance on a linear mapping assumption may also restrict the solution ' s ability to capture the complex dynamics and dependencies present in the glass manufacturing process .

[0033] Moreover, this known approach uses sequential optimisation and prioritisation of speci fic parameters , thus of fering a less comprehensive solution .

[0034] Furthermore , this solution alters the timing settings of the plungers , such as by delaying or advancing the lower dead point . It has been veri fied that altering the timing settings in real-time during the functioning of the machine can signi ficantly af fect the shape of the glass gobs and implies modi fying the mechanical setups during the functioning of the gob forming system .

[0035] Summary of the Invention

[0036] The aim of the present invention is to provide a gob forming system, an individual section machine , a method for producing the glass gobs and a related computer program product that overcome the issues mentioned above .

[0037] According to the present invention, a gob forming system, an individual section machine , a method for producing the glass gobs and a related computer program product are provided, as defined in the annexed claims .

[0038] Brief Description of the Drawings

[0039] For a better understanding of the present invention, preferred embodiments thereof are now described, purely by way of non-limiting examples and with reference to the attached drawings , wherein : - Figure 1 is a block diagram schematically showing an individual section machine for producing glass containers ;

[0040] Figure 2 is a schematical diagram showing a gob forming system of the individual section machine of Figure 1 ; and

[0041] - Figure 3 is a block diagram schematically showing a method for producing glass gobs , implemented by the gob forming system of Figure 2 .

[0042] Description of Embodiments of the Invention

[0043] In the following, elements common to the di f ferent embodiments have been indicated with the same reference numbers .

[0044] Figure 1 shows an individual section ( IS ) machine 10 for forming glass containers such as glass bottles .

[0045] The individual section machine 10 ( also referred to herein as "machine 10" ) is used in the hot end part of the glass container manufacturing process , as previously described .

[0046] The machine 10 comprises a molten glass supply 12 , a gob forming system 14 , a gob distribution arrangement 16 and plurality of individual sections 18a, 18b .

[0047] Figure 1 exemplarily shows two individual sections 18a, 18b, however it is evident that the machine 10 analogously can have a di f ferent number of individual sections ( e . g . , three or more ) .

[0048] Each individual section ( also referred to herein as " section" ) can have one or more blank moulds 18 ' and one or more blow moulds 18" . Figure 1 exemplarily shows one blank mould 18 ' and one blow mould 18" for each section 18a, 18b, however it is evident that the number of blank moulds 18 ' and blow moulds 18" can be different (e.g., two blank / blow moulds each section 18a, 18b, or more) .

[0049] In particular, the structure and functioning of the molten glass supply 12, the gob distribution arrangement 16 and the sections 18a, 18b is known and, for example, is already described in the present prior art section.

[0050] During use, the molten glass supply 12 generates molten glass 13 starting from input batched materials 11 and supplies the molten glass 13 to the gob forming system 14 that converts the molten glass 13 into glass gobs 15 (also referred to herein as "gobs 15") . The gobs 15 are distributed to the sections 18a, 18b by means of the gob distribution arrangement 16. Then, the sections 18a, 18b form the glass containers (here indicated with the reference 19) using the received glass gobs 15. This is iteratively repeated until all the desired glass containers 19 are formed.

[0051] In detail, the gobs 15 can be distributed to the sections 18a, 18b in succession (e.g., firstly to section 18a and then to section 18b) or they can be distributed according to other strategies, for example depending on the variations that the gobs 15 should have to obtain different glass containers 19 through the different sections 18a, 18b.

[0052] In particular, the glass containers 19 produced in the different sections 18a, 18b can have same physical properties (e.g., same length, weight, shape, etc.) or different physical properties among them. This second option allows to form a plurality of containers having different features by means of a single individual section machine 10 and during a single machine cycle (i.e., substantially simultaneously) , without the need of changing the setup of the machine 10. More in detail , the gobs 15 can be sent sequentially to the sections 18a , 18b, so that the glass containers 19 are not actually formed at the same time . In such a situation, the machine 10 may be continuously forming, in a cyclical way, a succession of glass containers 19 having di f ferent features . In other words , each section 18a, 18b forms glass containers 19 having substantially the same physical parameters but di f ferent from those of the glass containers 19 produced by the other section 18a, 18b , so that the overall machine 10 may output glass containers 19 having di f ferent physical features .

[0053] The gobs 15 produced during a cycle of the machine 10 ( i . e . , considering both the section 18a and the section 18b ) may be referred to herein as a set of gobs 15 .

[0054] Figure 2 shows in detail the gob forming system 14 , also referred to as a system for producing gobs .

[0055] The gob forming system 14 includes a gob feeder apparatus 30 , a shearing mechanism 32 , a gob forming control unit 34 and a gob sensing apparatus 36 .

[0056] In particular, the gob sensing apparatus 36 is operatively coupled to the gob feeder apparatus 30 and / or the shearing mechanism 32 to measure parameters of the gobs 15 ( also called in the following " gob parameters" or "gob features" ) , and the gob forming control unit 34 is coupled to the gob feeder apparatus 30 and the shearing mechanism 32 , to control them, and to the gob sensing apparatus 36 , to acquire the measured gob parameters .

[0057] During use , the gob feeder apparatus 30 receives from the molten glass supply 12 the molten glass flow and generates the glass streams 45 that are then cut by the shearing mechanism 32 to obtain the gobs 15 . Moreover, the gob sensing apparatus 36 detects the gob parameters , as better described in the following, and the gob forming control unit 34 acquires these gob parameters and accordingly controls the gob feeder apparatus 30 and the shearing mechanism 32 , thus also the formation of the gobs 15 .

[0058] As shown in Figure 2 , the gob feeder apparatus 30 comprises a refractory spout bowl 40 having a plurality of discharge ori fice holes 42 located in the bottom of the spout bowl 40 .

[0059] The number of discharge ori fice holes 42 is correlated to the number of moulds ( in particular, the blank moulds 18 ' and blow moulds 18" ) of each section 18a, 18b . In particular, the number of discharge ori fice holes 42 is equal to the number of blank moulds 18 ' and to the number of blow moulds 18" of each section 18a, 18b . Figure 2 exemplarily shows two discharge ori fice holes 42 corresponding to the case of the machine 10 having two blank moulds 18 ' and two blow moulds 18" in each section 18a, 18b . However, a di f ferent number of discharge ori fice holes 42 could be analogously considered ( e . g . , more than two ) .

[0060] The spout bowl 40 of the gob feeder apparatus 30 contains the molten glass 13 received from the molten glass supply 12 .

[0061] The gob feeder apparatus 30 further comprises a feeder tube 44 that extends into the spout bowl 40 . In particular, the feeder tube 44 is lowered into the spout bowl 40 to move towards , as well as away from, the discharge ori fice holes 42 . Moreover, the feeder tube 44 can also rotate axially with respect to the spout bowl 40 . These movements of the feeder tube 44 are executed by means of a feeder drive arrangement 46 of the gob feeder apparatus 30 .

[0062] The feeder drive arrangement 46 is configured to control the position of the feeder tube 44 relative to the bottom of the spout bowl 40 , i . e . to control the flow of molten glass 13 into the feeder tube 44 . In particular, the feeder drive arrangement 46 is coupled to the gob forming control unit 34 and is configured to receive a feeder tube control signal from the gob forming control unit 34 and to control the position of the feeder tube 44 based on the feeder tube control signal . Thus , the feeder drive arrangement 46 can comprise actuators (not shown) that are electrically controllable to vary the position of the feeder tube 44 with respect to the spout bowl 40 .

[0063] The gob feeder apparatus 30 further comprises a plurality of plunger needles 50 that extend within the feeder tube 44 . Each plunger needle 50 extends respectively above one of the discharge ori fice holes 42 . The plunger needles 50 are moved downwardly and upwardly in the feeder tube 44 by means of a plunger drive arrangement 48 of the gob feeder apparatus 30 , in order to respectively advance and retract the plunger needles 50 towards and away from the discharge ori fice holes 42 . These reciprocating movements of the plunger needles 50 selectively advance the molten glass 13 through the discharge ori fice holes 42 to generate the molten glass streams 45 .

[0064] The plunger drive arrangement 48 is configured to control the position of the plunger needles 50 relative to the bottom of the spout bowl 40 , i . e . to control the flow of molten glass 13 through the discharge ori fice holes 42 . In particular, the plunger drive arrangement 48 is coupled to the gob forming control unit 34 and is configured to receive a plunger control signal from the gob forming control unit 34 and to control the position of the plunger needles 50 based on the plunger control signal . Thus , the plunger drive arrangement 48 can comprise actuators (not shown) that are electrically controllable to vary the position of the plunger needles 50 with respect to the spout bowl 40 .

[0065] In further detail , the plunger drive arrangement 48 can control the plunger needles 50 either simultaneously or independently from each other . In other words , in the first case the movements of the plunger needles 50 are simultaneous and the same ( e . g . , in terms of speed and direction ) , whereas in the second case the movements of the plunger needles 50 are di f ferent and / or asynchronous .

[0066] The motion profile of the plunger needles 50 generated by the plunger drive arrangement 48 is referred to in the following as a cam . The cam profile thus is indicative of the movement and speed of the plunger needles 50 and of the dead points of the plunger needles 50 ( i . e . , of the starting and finishing positions of the plunger needles 50 during their reciprocating movements ) . The cam is thus controlled by the plunger control signal received from the gob forming control unit 34 . Since it is reciprocal motion, each cycle of the plunger needles 50 can be correlated to a 360-degree cycle , which simulates prior rotating mechanical cams used to convert rotational motion into linear motion .

[0067] The shearing mechanism 32 comprises a plurality of sets of shear blades 52 ( one set for each molten glass stream 45 ) and a shear drive arrangement 54 . For example, each set of shear blades 52 can be in the form of a reciprocating pair of parallel shear blades that are located below the spout bowl 40 and have opposed engageable cutting edges that are located below the respective discharge orifice hole 42. When the shear blades 52 of a set are reciprocated, they cut the molten glass stream 45 coming from the respective discharge orifice holes 42, thus generating a glass gob 15. However, this action can be executed also by means of other forms of the sets of shear blades 52 (e.g., through a single shear blade) .

[0068] The shear blades 52 are actuated by the shear drive arrangement 54 that is configured to control the relative positions of the shear blades 52, i.e. to control the timing of the cutting of the molten glass streams 45. In particular, the shear drive arrangement 54 is coupled to the gob forming control unit 34 and is configured to receive a shear control signal from the gob forming control unit 34 and to control the position of the shear blades 52 based on the shear control signal. Thus, the shear drive arrangement 54 can comprise actuators (not shown) that are electrically controllable to reciprocate the shear blades 52.

[0069] In further detail, the shear drive arrangement 54 can control the different sets of shear blades 52 either simultaneously or independently from each other. In other words, in the first case the movements of the shear blades 52 of each set are simultaneous and the same (e.g., in terms of speed and direction) , whereas in the second case the movements of the shear blades 52 are different and / or asynchronous between the different sets.

[0070] As previously stated, the gob sensing apparatus 36 is configured to measure one or more gob parameters.

[0071] The gob parameters are parameters indicative of features of the gobs 15. In particular but not exclusively, the gob parameters include physical parameters of the gobs 15.

[0072] In detail, the one or more gob parameters include the mass of the gobs 15 (in the following referred to as "gob mass") .

[0073] However, the gob parameters may also include one or more of the following parameters: the length of the gobs 15 (in the following referred to as "gob length") ; the shape of the gobs 15 (in the following referred to as "gob shape") , for example a 2D or 3D shape; and the diameter of the gobs 15 (in the following referred to as "gob diameter") .

[0074] These parameters can be better defined considering that the gobs 15 have substantially cylindrical shape. In fact, the cylindrically shaped molten glass streams 45 exiting the discharge orifice holes 42 are cut by the shearing mechanism 32 along cross-sections of the molten glass streams 45, thus substantially resulting in cylindrically shaped gobs 15.

[0075] In detail, each gob 15 has upper and bottom surfaces which substantially correspond to the circular bases of the cylinder and thus to the cross-sections of the respective molten glass stream 45 along which the cuts occur. The diameter of the substantially circular upper and bottom surfaces (i.e., of cross-sections of the gob 15) is the gob diameter. Moreover, the gob length corresponds to the height of each gob 15, i.e. to the longitudinal length of the cylinder and thus to the relative distance between the upper and bottom surfaces. As evident , these definitions properly apply to the case of a perfectly-shaped cylinder whereas the shape of the gobs 15 is only an approximations of this shape ( e . g . , due to deformations of the molten glass streams 45 that cause slight longitudinal bending of the gobs 15 and to the transversal compression exerted by the shearing mechanism 32 on the molten glass streams 45 that cause the deformation of the upper and bottom surfaces towards shapes such as ogival ellipses ) ; however, any analogous definition can be considered i f systematically applied, as obvious . For example , the gob length can be calculated as an average of the longitudinal distances between the upper and bottom surfaces measured on the lateral surface of the gob 15 , or as the minimum / maximum one of these distances . Analogously, the gob diameter can be defined as an equivalent diameter of the upper / bottom surface of the gob 15 .

[0076] The gob parameters may also include additional parameters , such as manufacturing process parameters , that impact on the physical properties of the gobs 15 . For example , these additional parameters can be the temperature and / or the viscosity of the gobs 15 or of the molten glass 13 .

[0077] In view of the gob parameters to be measured, the gob sensing apparatus 36 comprises respective sensors of per se known type .

[0078] In particular, the gob sensing apparatus 36 comprises a mass sensor for detecting the gob mass .

[0079] The mass sensor can comprise one or more image acquisition devices 56 ( e . g . , cameras ) that are configured to acquire images of the gobs 15 and to determine the gob mass based on these images , in a per se known way .

[0080] For example and as exemplarily shown in Figure 2 , there is one image acquisition device 56 for each set of shear blades 52 , so that the cutting of all the molten glass streams 45 is monitored . The image acquisition devices 56 are placed at the shear blades 52 and face the space below the shear blades 52 so that the gobs 15 can be imaged after being cut by the shear blades 52 . The acquired images are then used to determine a volume of the gobs 15 in a per se known way and thus to determine the corresponding mass , considering the known density of the glass . However, it is evident that also a single image acquisition device 56 ( e . g . , a 360 ° camera ) can be suf ficient to image the gobs 15 coming from all the molten glass streams 45 , if correctly positioned so as to observe the traj ectory of all the gobs 15 produced by the gob feeder apparatus 30 .

[0081] According to a di f ferent example (not shown) , there is a plurality of image acquisition devices 56 ( e . g . , a camera linear array or matrix ) for all the sets of shear blades 52 , in particular two image acquisition devices 56 configured to image the gobs 15 generated starting from all the molten glass streams 45 . The image acquisition devices 56 are angulated with respect to one another, so that they have respective observation directions that are inclined with respect to each other and directed towards the gobs 15 . Therefore , the image acquisition devices 56 can capture images of the gobs 15 from several points of view, thus allowing a 3D reconstruction of the gobs 15 or at least a more detailed evaluation of the gobs 15 ( i . e . , not only an evaluation from a single 2D image obtained from a single point of view) .

[0082] The image acquisition devices 56 can also be configured to measure at least one among the gob length, the gob shape and the gob diameter based on the acquired images of the gobs 15 .

[0083] In particular and considering the exemplary embodiment of Figure 2 wherein one image acquisition device 56 is present for each set of shear blades 52 , it is possible to determine the gob length, the gob shape ( in detail , the 2D shape ) and the gob diameter starting from the acquired image .

[0084] On the other hand, considering the exemplary embodiment wherein more than one image acquisition devices 56 is present for each set of shear blades 52 , with respect to the previously described case it is possible to determine more accurately the gob length, the gob shape ( in detail , it is also possible to determine the 3D shape ) and the gob diameter, by simultaneously taking into account the images acquired by all the image acquisition devices 56 of the same set of shear blades 52 .

[0085] Moreover, the gob sensing apparatus 36 can also comprise additional sensors for measuring the above-mentioned additional parameters . For example , the gob sensing apparatus 36 can also comprise one or more of the following sensors : one or more temperature sensors (not shown, such as an IR sensor with a field of view pointing towards the traj ectory of the falling gobs 15 or towards the inside of the spout bowl 40 ) for measuring the temperature of the gobs 15 or of the molten glass 13 ; and one or more viscosity sensors (not shown, for example placed in the spout bowl 40 or in the feeder tube 44 ) for measuring the viscosity of the gobs 15 or of the molten glass 13.

[0086] The gob forming control unit 34 is an electronic control unit (e.g., a processor, a microprocessor, etc.) configured to acquire and process data, as better described in the following, in order to control the gob forming system 14 for producing the gobs 15.

[0087] During use, the gob forming control unit 34 acquires the gob parameters from the gob sensing apparatus 36 and controls accordingly the formation of the gobs 15. In particular, the gob forming control unit 34 controls the plunger needles 50, the feeder tube 44, the shear blades 52 and any other relevant setting of the gob forming system 14 based on the gob parameters. For examples, the gob forming control unit 34 generates the feeder tube control signal, the plunger control signal, the shear control signal and any other control signal useful for the setting of the gob forming system 14 based on the gob parameters.

[0088] In fact, the settings of the gob forming system 14 determine the physical features of the gobs 15 produced according to these settings. In general, these settings comprise control parameters of the gob forming system 14. For example, the above-mentioned feeder tube control signal, plunger control signal and shear control signal are examples of signals indicative of these control parameters.

[0089] In particular, the control parameters (or setting parameters) comprise one or more of the following parameters: the movement profile of the plunger needles 50 (i.e., the cam profile) ; the position and / or the rotational speed of the feeder tube 44 relative to the bottom of the spout bowl 40; and the phase of the shear blades 52. According to an embodiment , the control parameters comprise at least the positions of the plunger needles 50 and may also comprise one or more of the other mentioned parameters .

[0090] During use , the gob forming control unit 34 implements a model and an optimiser in order to control the gob forming system 14 , as better discussed in the following with reference to Figure 3 .

[0091] In detail , the model comprises an arti ficial intelligence architecture able to determine output data starting from input data provided to the model , after a suitable training . Due to this association between input and output parameters generated by the trained model , the model can be defined as a data-driven one . In particular, this corresponds to relying on the use of a training to perform this data association rather than, as commonly done in the prior art solution, using physics-based approaches to calculate the output data based on the input data ( e . g . , to calculate the output data by processing the input data by means of predefined mathematical functions that have a clear physical explanation, for example by using linear matrices multiplied to the input data or representing the system using di f ferential equations in order to calculate the output data ) .

[0092] In particular, during a training phase the model receives as input a training dataset comprising a plurality of sets of control parameters and a corresponding plurality of sets of gob parameters . Each set of control parameters comprises the values of the control parameters that can be used in the gob forming system 14 to produce a gob 15 having some predefined and target physical properties ( i . e . , the values of the gob parameters of the set of gob parameters associated to the considered set of control parameters ) . Therefore , the model can be trained in a supervised manner to be able to associate to each set of control parameters a corresponding set of gob parameters .

[0093] In further detail , the model is a non-linear model .

[0094] According to an embodiment of the invention, the model comprises a Neural Network feeder model , such as a multilayer perceptron (MLP ) model .

[0095] In detail , the neural network comprises multiple layers , starting with an input layer that receives the control parameters ( e . g . , data pertaining to cam profile , height of the feeder tube 44 and the phase of the shear blades 52 ) . Following the input layer, one to three hidden layers may be equipped with activation functions working together to process the input data, extracting complex patterns and relationships that are not readily discernible through traditional analytical methods . In fact , a signi ficant feature of this model is its ability to adapt dynamically, undergoing periodical adj ustments to fine-tune its predictions . This is facilitated through a feedback loop that can integrate real-time data from the manufacturing floor ( e . g . , some of the control parameters ) , allowing the network to learn and adapt to the ever-changing dynamics of the production process . The continuous adaptation ensures that the model remains attuned to the variations and dri fts in the manufacturing process , optimising its predictions over time to achieve desired output parameters with higher accuracy .

[0096] In terms of architecture , the neural network can adopt the MLP structure , comprising a series of densely connected layers that facilitate deep learning . The input layer is designed to accommodate a wide array of features , which are then processed through hidden layers equipped with nonlinear activation functions to capture the complex relationships in the data . The output layer, compri sing nodes representing the weight , length and shape of the glass gobs 15 , is optimised through backpropagation algorithms , minimising the error between predicted and actual values and thus enhancing the model predictive accuracy . Moreover, the accuracy of the model is improved through the integration of real-time data and periodic updating .

[0097] According to a di f ferent embodiment , the model comprises a Stochastic Gradient Descent ( SGD) model .

[0098] SGD is a variation of the Gradient Descent ( GD) algorithm that addresses computational ef ficiency and scalability concerns when dealing with large datasets . Unlike GD, which computes the gradient of the cost function using the entire training dataset , SGD randomly selects individual training examples during each iteration and performs gradient updates based on these examples . This approach introduces a level of stochasticity in the parameter updates , allowing SGD to make progress towards the optimal solution even with a subset of the data . More detai ls about the SGD can be found for example at the link sklearn . linear_model . SGDRegressor — scikit-learn 1 . 3 . 1 documentation .

[0099] In each iteration, SGD randomly selects a training example and computes the gradients using that example . The gradients are then used to update the model parameters . This process continues until a predefined stopping criterion is met , such as a maximum number of iterations or convergence of the cost function .

[0100] SGD allows for online learning, where new training examples can be incorporated into the model without requiring a complete re-training process .

[0101] However, due to the random selection of training examples , the parameter updates in SGD may have higher variance compared to GD . To mitigate this issue , techniques such as mini-batch SGD can be used, where a small random batch of training examples is used to compute the gradients and update the parameters . This approach strikes a balance between the computational ef ficiency of SGD and the stability of GD .

[0102] The optimiser comprises arti ficial intelligence tools that , using the trained model and considering a target solution, allow to find an optimal solution among a plurality of candidate solutions ( i . e . , a solution that approaches the target solution better than the other candidate solutions ) . In other words , the optimiser is a meta-heuristic optimiser .

[0103] The working of the optimiser is better described in the following . However, in general the optimiser receives as an input target gob parameters ( i . e . , the target physical features of the gob 15 that should be produced) and, using the trained model , finds which candidate set of control parameters leads to a correspondent candidate set of gob parameters that better approaches the target gob parameters , so that the gob forming system 14 can be controlled through this optimal set of control parameters to obtain a respective gob 15 that is s imilar to the target one . In view of the fact that the meta-heuristic optimiser generates candidate solutions while exploring the space to find the most appropriate candidate solution, the optimiser can be defined as a data-driven optimiser .

[0104] According to an embodiment , the optimiser comprises a Nelder-Mead optimiser .

[0105] The Nelder-Mead algorithm, also known as the downhill simplex algorithm, is a derivative- free optimisation technique that can be adapted for the problem of glass manufacturing processes . It is particularly suitable for problems where the obj ective function is not di f ferentiable or where the gradient information is not readily available .

[0106] In the context of controlling gob parameters , the Nelder-Mead optimiser aims to find the optimal set of control variables that minimise the discrepancy between the target gob parameters and the predicted ( or candidate ) gob parameters . This is achieved by iteratively adj usting the candidate control parameters and evaluating the resultant predicted gob parameters , to converge towards the optimal solution .

[0107] Mathematically, the Nelder-Mead algorithm operates on a simplex, which is a set of n+ 1 points in an n-dimensional space . Each point represents a candidate solution, and the algorithm updates the simplex iteratively to search for the optimal solution .

[0108] Considering x± as the representation of the ithpoint in the simplex, where i=l , 2 , n+ 1 , the goal is to minimise the obj ective function f (x ) , which represents the discrepancy between the desired ( target ) and predicted ( candidate ) gob parameters ( e . g . , sum of squared errors , for instance ) . In detail, the Nelder-Mead algorithm can proceed through the following steps: initialisation (generate an initial simplex by selecting n+1 points in the search space, wherein each point corresponds to a set of control variables) ; evaluation (evaluate the objective function f (x) for each point in the simplex, representing the discrepancy between the desired and predicted gob weight and length values) ; sorting (sort the simplex based on the objective function values; the point with the best objective function value represents the best solution found so far) ; reflection (compute the reflection point by reflecting the worst point across the centroid of the remaining points in the simplex, then evaluate the objective function for the reflection point) ; expansion (if the objective function value of the reflection point is better than the best point in the simplex, compute the expansion point by further extending from the reflection point, then evaluate the objective function for the expansion point) ; contraction (if the objective function value of the reflection point is worse than the second worst point in the simplex, perform a contraction, then compute the contraction point by moving towards the centroid from the reflection point and evaluate the objective function for the contraction point) ; shrink (if none of the above conditions are met, perform a shrink operation, then compute new points by shrinking the simplex towards the best point and evaluate the objective function for the new points) ; update (update the simplex based on the evaluation results, ensuring that the best point is always retained) ; and termination (repeat steps reflection to update until a termination criterion is met, such as reaching a maximum number of iterations or achieving a desired level of convergence ) .

[0109] By applying the Nelder-Mead optimi ser to the control of gob weight or the weight and length, it is possible to iteratively adj ust the candidate control parameters to minimise the discrepancy between the desired and predicted values .

[0110] The Nelder-Mead optimiser ef ficiently explores the search space by iteratively updating the simplex based on reflection, expansion, contraction, and shrink operations . Through these iterations , the algorithm converges towards the optimal set of control variables that minimise the discrepancy between the desired and predicted gob weight and length values .

[0111] In summary, the Nelder-Mead optimiser provides an ef fective approach for the control of the features in the glass manufacturing process . By iteratively updating the simplex and minimising the obj ective function, the algorithm enables the optimisation of the control parameters to approximate the desired gob parameters , resulting in improved gob quality and reduced variations in gob features .

[0112] According to a di f ferent embodiment , the optimiser comprises a Simulated Annealing algorithm .

[0113] The Simulated Annealing algorithm is a meta-heuristic algorithm that emulates the anneal ing process in metallurgy to search for the global optimum in a given search space . It can be ef fectively applied to control the features of the glass gobs 15 . The algorithm begins with an initial solution and iteratively explores the search space by considering neighbouring solutions . The quality of a solution is evaluated using an obj ective function that measures the discrepancy between the desired gob parameters and the predicted gob parameters inferenced by the model .

[0114] At each iteration, the Simulated Annealing algorithm accepts or rej ects a neighbouring solution based on a probabilistic acceptance criterion . This criterion allows the algorithm to occasionally accept worse solutions , which helps in escaping local optima and exploring di f ferent regions of the search space . The probability of accepting a worse solution is determined by the current temperature and the di f ference in obj ective function values . As the temperature decreases during the optimisation process , the acceptance probability reduces , biasing the search towards better solutions . The cooling schedule controls the rate at which the temperature decreases and can be linear or nonlinear, depending on the problem. The Simulated Annealing process continues until a stopping criterion is met, such as reaching a maximum number of iterations or achieving a desired level of convergence .

[0115] By iteratively exploring the search space , considering neighbouring solutions , and gradually reducing the temperature , the algorithm is able to ef fectively search for the global optimum, leading to improved control of gob weight and length .

[0116] In the present context , the Simulated Annealing algorithm can be integrated into the overall control system . The obj ective function could be defined to minimise the discrepancy ( sum of squared errors , for instance ) between the target gob parameters and the predicted gob parameters obtained based on the candidate control parameters . The algorithm can iteratively adj ust the control parameters using the acceptance probability and the cooling schedule to guide the search towards optimal solutions .

[0117] According to a di f ferent embodiment , the optimiser comprises an adaptive multi-obj ective simulated annealing (AMOSA) algorithm .

[0118] In addition to the Simulated Annealing optimiser, the multi-ob ective version known as Adaptive Multi-Obj ective Simulated Annealing (AMOSA) can be employed in the glass manufacturing process . AMOSA extends the Simulated Annealing algorithm by considering multiple obj ective functions simultaneously, allowing for a more comprehensive optimisation on several parameters at once .

[0119] In the context of controlling the gob features , the AMOSA optimiser aims to find the optimal set of control variables that achieve desired gob parameters for each section . This is achieved by simultaneously optimising multiple obj ective functions , such as minimising the discrepancy between the desired and predicted gob parameters for each section .

[0120] The AMOSA algorithm iteratively explores the search space by generating neighbouring solutions and evaluating their obj ective function values . Similar to Simulated Annealing, the acceptance criterion is used to determine whether to accept or rej ect a neighbouring solution based on the multi-obj ective obj ective function values .

[0121] The acceptance probability for a worse solution in AMOSA is calculated using the Hypervolume Indicator (HVI ) , which quanti fies the improvement in the Pareto front ( a set of non-dominated solutions ) achieved by the neighbouring solution . The probability of accepting a worse solution is determined by the current temperature and the HVI value .

[0122] By considering multiple obj ective functions and using the HVI as the acceptance criterion, the AMOSA algorithm can guide the search towards the Pareto-optimal solutions , which represent the trade-of f between the gob parameters for each section . This allows for a more comprehensive optimisation approach that considers the speci fic requirements of each section .

[0123] Integrating the AMOSA optimiser into the control system enables the simultaneous optimisation of all the gob parameters . The obj ective functions are defined to minimise the discrepancies between the desired and predicted gob parameters for each section . The algorithm iteratively adj usts the control parameters based on the acceptance probability calculated using the HVI and the cooling schedule .

[0124] According to a di f ferent embodiment , the optimiser comprises a Genetic Algorithm .

[0125] The Genetic Algorithm is a powerful optimisation technique inspired by the principles of natural selection and genetics . It can be applied to the control of gob features in the glass manufacturing process , providing an ef ficient way to search for optimal control parameters .

[0126] In the context of controlling gob features , the Genetic Algorithm works by evolving a population of candidate solutions represented as individuals . Each individual corresponds to a set of control parameters . The algorithm applies genetic operators , including selection, crossover, and mutation, to iteratively generate new generations of individuals that better match the desired gob parameters . Algorithmically, the Genetic Algorithm can operate as follows : initialisation ( generate an initial population of individuals randomly, wherein each individual represents a set of control variables that determine gob weight , length and shape ) ; evaluation ( evaluate the fitness , i . e . , cost function, of each individual in the population using an obj ective function that measures the discrepancy between the desired and predicted gob weight and length values ) ; selection ( select individuals from the population based on their fitness ; individuals with higher fitness have a higher probability of being selected for reproduction) ; crossover (perform crossover operations between selected individuals to create of fspring ; the crossover involves exchanging genetic information between parents to produce new solutions that combine their features ) ; mutation ( introduce random changes in the of fspring' s genetic information to maintain diversity and explore new regions of the search space ; mutation helps prevent premature convergence to suboptimal solutions ) ; replacement ( replace some individuals in the current population with the newly created of fspring ; the replacement strategy can be based on factors like fitness or age ) ; termination ( repeat steps of evaluation to replacement until a termination criterion is met , such as reaching a maximum number of generations or achieving a desired level of convergence ) .

[0127] The Genetic Algorithm iteratively evolves the population of individuals to optimise the control parameters , through the process of selection, crossover, and mutation . The fitness of each individual is evaluated based on the obj ective function, which measures the discrepancy between the desired and predicted gob parameters.

[0128] In summary, the Genetic Algorithm offers an effective approach for the control of the gob properties in the glass manufacturing process. By evolving the population of individuals and applying genetic operators, the algorithm enables the optimisation of the control parameters to approximate the desired gob parameters, resulting in improved gob quality and reduced variations in gob features.

[0129] In particular, the Genetic Algorithm can be a Nondominated Sorting Genetic Algorithm III (NSGA-III) . This algorithm is the multi-objective version of the Genetic Algorithm.

[0130] The NSGA-III extends the basic Genetic Algorithm by maintaining a diverse set of solutions that represents the trade-off between different objectives.

[0131] Algorithmically, the steps are similar to what was previously described. However, the NSGA-III considers the conflicting objectives of minimising the discrepancy between desired and predicted gob parameters for each section. This is formulated as a multi-objective optimisation problem.

[0132] The NSGA-III algorithm incorporates non-dominated sorting and environmental selection to maintain a diverse set of solutions in the population. By exploring the tradeoff between the different objectives, it enables the identification of Pareto-optimal solutions that represent the best compromises between the conflicting objectives. This multi-objective approach provides decision-makers with a range of trade-off solutions to choose from, based on their specific preferences and requirements.

[0133] During use, the gob forming control unit 34 controls the gob forming system 14 according to the following method for producing gobs 15, in order to control the production of the gobs 15.

[0134] The method is schematically shown in Figure 3, with the reference number 100. The method 100 is an iterative (i.e., cyclical) method and is implemented by the gob forming control unit 34.

[0135] With reference to Figure 3, it is exemplarily described an updating iteration of the method 100, i.e. an iteration that leads to the updating of the control parameters and to the consequent control of the gob forming system 14.

[0136] As better described in the following, each updating iteration may comprise a plurality of sub-iterations (also called machine cycles) wherein the gob forming system 14 is controlled using the lastly updated control parameters (i.e., the control parameters obtained in the immediately precedent updating iteration) .

[0137] With reference to Figure 3, at a step S10 of the method 100, the gob forming system 14 is initialised. This step is not executed at each updating iteration but it is executed only once, upon starting of the method 100. For example, it is executed when the gob forming system 14 is switched on for starting the production of gobs 15.

[0138] In particular, at step S10 the model, the optimiser and the general setup of the gob forming system 14 are initialised .

[0139] Concerning the model, the initialisation leads to the selection of model initialising parameters.

[0140] For example, the model initialising parameters can comprise the architecture of the model to be used. The architecture can be selected among a plurality of alternatives (e.g., the previously mentioned architectures) or can be fixed and predefined. In the first case, the architecture can be chosen either automatically, e.g. based on the specific task to be performed by the gob forming system 14 and on the target gob properties, or manually, e.g. by a user operating or supervising the gob forming system 14. In the second case, the architecture is chosen during design of the gob forming system 14 and cannot be changed .

[0141] The model initialising parameters can also comprise a training acquisition number of sub-iterations, i.e. a parameter indicative of the number of sub-iterations required for collecting the data required for a new training of the model, as better described in the following. During these sub-iterations, a sufficient amount of data for a new training of the model is acquired and stored so that it can be used for the new training. The selection of this parameter depends on factors such as the dynamics of the glass manufacturing process and the desired accuracy of the model. For example, the training acquisition number can be equal to 10.

[0142] The model initialising parameters can also comprise the selection of the gob parameters to be considered and modelled, among the previously described choices available. In fact, depending on the specific task to be executed it is possible that some of the gob parameters can be of interest whereas others not. In particular, the selected gob parameters comprise the gob mass and, for example, also the gob length. The model initialising parameters can also comprise the selection of the control parameters to be considered and used in the model, among the previously described choices available. In fact, depending on the specific task to be executed it is possible that some of the control parameters can be of interest whereas others not. For example, the selected control parameters can comprise the movement profile of the plunger needles 50 and the position of the feeder tube 44 relative to the bottom of the spout bowl 40.

[0143] The model initialising parameters can also comprise the model training approach, in particular with respect to the trainings successive to the first one (i.e., the retrainings) . In particular, there are two options to re-train the model: full re-training approach and model updating approach. In detail, the full re-training approach creates a new instance of the model at each re-training (i.e., at each updating iteration) and is compatible with all previously described models, whereas the model updating approach exploits continual training to update the parameters of the original model based on the new incoming data and is only compatible with some of the previously described models (e.g., the Stochastic Gradient Descent) .

[0144] Concerning the optimiser, the initialisation leads to the selection of optimiser initialising parameters.

[0145] For example, the optimiser initialising parameters can comprise the architecture of the optimiser to be used. The architecture can be selected among a plurality of alternatives (e.g., the previously mentioned architectures) or can be fixed and predefined. The selection can be made analogously to the model. The optimiser initialising parameters can also comprise the selection of the gob parameters to be considered and modelled . According to an exemplary and non-limiting embodiment , the selected gob parameters of the optimiser are the same with respect to the ones of the model .

[0146] The optimiser initialising parameters can also comprise the selection of the control parameters to be considered and used in the model . According to an exemplary and non-limiting embodiment , the selected control parameters of the optimiser are the same with respect to the ones of the model .

[0147] The optimiser initialising parameters can also comprise target gob parameters , i . e . target values for each of the selected gob parameters . In detail , the target gob parameters can also be di f ferent for the di f ferent individual sections 18a, 18b that are considered, i f di f ferent features of the gobs 15 are desired for the individual sections 18a, 18b . For example , in the exemplary and non-limiting case wherein the selected gob parameters are the gob mass and the gob length and wherein two individual sections 18a, 18b are present , the target gob parameters can comprise two vectors , one for the gob mass values and the other for the gob length values , each having two instances corresponding respectively to the target values of the two individual sections 18a, 18b . For example , the target gob parameters are received as input by the gob forming system 14 , e . g . are manually provided by the user .

[0148] The optimiser initialising parameters can also comprise an optimisation criterion ( also known as cost function) . In the context of the present meta-heuristic and data-driven optimisation algorithms , the cost function serves as an evaluative measure to quanti fy the discrepancy between the gob parameters predicted by the model and the target gob parameters . This function allows the optimiser to assess the quality of the di f ferent sets of feeder parameters , guiding the optimiser towards the optimal set that minimises the error . Speci fically, the cost function is formulated to take into account the gob parameters . Examples of optimisation criteria include : Sum of Squared Errors ( SSE ) ; Mean Squared Error (MSE ) ; Mean Absolute Error (MAE ) . In particular, SSE and MSE emphasise larger errors by squaring the individual discrepancies , making it sensitive to outliers but generally providing a more punishing error metric ; on the other hand, MAE gives a linear penalty to errors , which makes it less sensitive to outliers but can be easier to interpret and implement . By optimising such cost functions using metaheuristic techniques , the optimiser systematically explores the parameter space to find the most ef ficient settings to minimise the optimisation criterion, thereby achieving a more controlled and precise gob manufacturing process . The optimisation criterion can be selected analogously to what previously disclosed for the model architecture .

[0149] The optimiser initialising parameters can also comprise control parameter limits , i . e . ranges of variations or end values of some of the control parameters , beyond which the gob forming system 14 could not correctly work . This is useful to guide the optimi zer during the search of the optimal solution through the available space , since it prevents the optimiser from looking into solutions that could not be accepted by the gob forming system 14 due to physical constraints of the same . For example , the control parameter limits comprise cam profile movement limits (e.g., a lower dead point equal to 0 mm and a stroke maximum equal to 32 mm, measured with respect to respective zero positions for example defined by the user in a guided procedure that precedes the real-time use of the machine 10) and feeder tube height limits (e.g., a minimum height equal to 5 mm and a maximum height equal to 9 mm, measured with respect to respective zero positions for example defined in the guided procedure) .

[0150] The optimiser initialising parameters can also comprise additional specific optimiser parameters (e.g., maximum number of iterations, maximum temperature for Simulated Annealing, number of individuals and maximum generations in Genetic Algorithms, etc.) . The specific optimiser parameters can be selected analogously to what previously disclosed for the model architecture.

[0151] Concerning the setup of the gob forming system 14, the initialisation leads to the selection of system initialising parameters .

[0152] The system initialising parameters can comprise for example a re-training frequency (also called frequency of action) . The re-training frequency is a parameter indicative of the minimum number of sub-iterations required for a new training of the model, as better described in the following. In other words, the re-training frequency is the maximum frequency at which re-training can occur, i.e. is indicative of the minimum number of sub-iterations that can elapse between two consecutive re-trainings of the model.

[0153] According to an embodiment, the re-training frequency can be correlated to the training acquisition number and in particular can be the inverse of a number of sub-iterations which is greater than the training acquisition number . In fact , in this embodiment in order to be able to execute a new training it i s required firstly to collect and store the training data ( i . e . , to wait for a time defined by the training acquisition number ) and then to actually re-train the model based on the collected data . The re-training frequency can thus be correlated to the inverse of the total amount of sub-iterations required for this process , i . e . as the inverse of the time length of the updating iteration . Consequently, in this embodiment the multiplication of the re-training frequency by the cycle time ( i . e . , the time period required for performing a sub-iteration and thus for producing a set of gobs 15 ) is obviously greater than the time required by the optimiser to generate inferences .

[0154] According to a di f ferent embodiment , considered in the following for exemplary purposes , the re-training frequency is uncorrelated to the training acquisition number and in particular is indicative of the minimum number of subiterations required to re-train the model and use the optimiser when the training data are continuously stored in a buf fer, so that the training data for a new training are available immediately after the execution of the previous training . Therefore , in this case the re-training frequency is independent from the time required to acquire the training data and depends mainly on the speed for training the model and the speed for running the optimiser .

[0155] The selection of the re-training frequency depends on factors such as the dynamics of the glass manufacturing process and the desired accuracy o f the model . For example , in the second embodiment the re-training frequency can correspond to three sub-iterations .

[0156] The system initialising parameters can also comprise the storage option, indicative of the storage approach used to memorise the training data when collected . In particular, two options can be available : ephemeral approach and cumulative approach .

[0157] In the ephemeral approach, the training data that are collected during a training acquisition number of subiterations are used for the one or more training iterations occurring during said training acquisition number of subiterations and are then deleted . In other words , in this approach the training data used at a considered updating iteration are acquired during a training acquisition number of sub-iterations that are immediately precedent to the considered updating iteration, are stored in the buf fer for said training acquisition number o f sub-iterations and then are deleted from it ( i . e . , the buf fer stores information only about the last training acquisition number of subiterations prior to the considered updating iteration ) . Considering the exemplary values provided before , the training data collected at each sub-iteration can be used for example in four subsequent trainings of the model . In the following description, this approach is exemplarily considered .

[0158] On the other hand, in the cumulative approach, the training data that are collected are added to a cumulative persistent database that is used for re-training the model . This approach of fers a more general isable model with a better understanding of process dynamics and a wider search space ; however, the training data accumulation increases the training time . This restraint can be remedied through a slight variation of this approach, i . e . by limiting the maximum number of training data that can be stored in the persistent database ( e . g . , the training data corresponding to 100 sub-iterations , for instance ) . In this way, after reaching the maximum number of training data that can be stored in the database , each new incoming training datum substitutes the oldest training datum stored in the database so that the number of training data does not exceed this threshold and the training time remains limited . This variation is similar to the ephemeral approach, except for the number of data that the persistent database can store ( generally much higher than the number of data that the buf fer can store ) .

[0159] The system initialising parameters can also comprise any additional parameter of the gob forming system 14 that can be initially required to operate the gob forming system 14 . For example , an initialisation set of control parameters can be present , for controlling the gob forming system 14 at the first sub-iteration of the method 100 .

[0160] At a step S 12 of the method 100 , the gob forming system 14 is controlled in order to produce a set of gobs 15 .

[0161] In particular, during the first updating iteration the gob forming system 14 is controlled based on the initialisation set of control parameters , i . e . on a set of control parameters that is predefined ( for example that is chosen based on the target gob parameters ) , whereas during the subsequent updating iterations the gob forming system 14 is controlled at each sub-iteration based on the sets of control parameters previously obtained through the optimiser as the optimal control parameters for approximating the target gob parameters , i . e . based on the last set o f control parameters that has been obtained . Therefore , at each subiteration, the gob forming system 14 is controlled through a respective set of control parameters in order to produce a respective set of gobs 15 .

[0162] At a step S 14 of the method 100 , the gob forming control unit 34 acquires the sets of gob parameters measured by the gob sensing apparatus 36 during the sub-iterations .

[0163] In detail , a respective set of gob parameters is acquired for each set of gobs 15 produced by controlling the gob forming system 14 with a respective set of control parameters at a respective sub-iteration of the considered updating iteration .

[0164] Considering exemplarily the ephemeral approach as previously stated, each new set of control parameters and the respective set of gob parameters are stored in the buf fer and thus are part of the training dataset on which the training ( or re-training) of the model will be based at the next updating iteration . Otherwise , considering exemplarily the cumulative approach, these data are collected and stored into the cumulative persistent training database .

[0165] At a step S 16 of the method 100 , the model is trained ( i f considering the first updating iteration) or re-trained ( i f considering the subsequent updating iterations ) based on the training data acquired at the step S 14 . In particular, the re-training is performed at the re-training frequency .

[0166] As previously mentioned, the training or re-training ( i . e . , updating of the training) can be executed according to the ephemeral approach or the cumulative approach. By way of example and considering the use of the ephemeral approach, the re-training frequency corresponding to three subiterations and the training acquisition number being equal to ten sub-iterations, during each updating iteration the re-training is performed based on the buffer storing the sets of control parameters and gob parameters acquired at the ten sub-iterations immediately precedent to the considered updating iteration. For example, the training is stopped when a training stopping criterium is met (e.g., in terms of maximum epochs or sufficient accuracy of the model) .

[0167] At a step S18 of the method 100, an optimal set of control parameters is found through the use of the optimiser and based on the trained model.

[0168] In particular, the optimiser initially suggests a candidate solution (i.e., a candidate set of control parameters) based on either random selection or some heuristic techniques (this choice depends on the specific optimiser that is chosen, as known) . This candidate solution is then tested using the model trained in step S16, to evaluate its performance. In particular, this evaluation is performed by comparing the candidate set of gob parameters, obtained by the model processing the candidate set of control parameters received as input, with the target gob parameters through the specific optimisation criterion (e.g., the SSE cost function) . Both the target gob parameters and the optimisation criterion are among the inputs of the optimiser, as previously described. The optimisation criterion measures how well the candidate set of gob parameters meets the target gob parameters, such as by minimising the errors between each type of gob parameter . The optimiser then iterates over within the defined search space considering other candidate solutions ( determined according to the speci fic optimiser used, as mentioned before and as known) , evaluating them each time through the model and the associated optimisation criterion . For example , the search space is defined according to the control parameter limits , better discussed previously . This iterative process continues until the algorithm converges on the best solution, i . e . the optimal set of control parameters that provides the optimal set of gob parameters that better approximate the target gob parameters among the examined candidate sets of control parameters .

[0169] At a step S20 of the method 100 , the control parameters used for controlling the gob forming system 14 are updated based on the optimal set of control parameters . In particular, this optimal set replaces the previously stored control parameters and the considered updating iteration ends .

[0170] After step S20 , the method returns to step S 12 and thus the gob forming system 14 is controlled based on the updated set of control parameters in order to produce new sets of gobs 15 .

[0171] In particular, whereas the steps S 12 and S 14 refer to actions performed at each sub-iteration, the steps S 16-S20 are performed once for each updating iteration based on the training data previously acquired . Therefore , during the execution of steps S 16-S20 the gob forming system 14 can be continuously controlled based on the available control parameters , i . e . the set of control parameters determined at the previous updating iteration . This allows to keep on generating sets of gobs 15 even during the time period required by the model to train or re-train and by the optimiser to find the optimal set of control parameters . In this way the production of gobs 15 is not interrupted by the actions of steps S 16-S20 . The periodic re-training of the model and the periodic updating of the control parameters based on the re-trained model allows to obtain a dynamic response of the gob forming system 14 with respect to inevitable small changes in the manufacturing process that could af fect the quality of the gob production . For example , this can overcome changes in environmental factors such as ambient temperature or molten glass viscosity . In view of this , the model can be considered a dynamical one .

[0172] The method 100 is implemented in the gob forming system 14 by means of a correspondent computer program product stored in the gob forming system 14 .

[0173] From what has been described and illustrated previously, the advantages of the present invention are evident .

[0174] The present solution allows to accurately control the physical and process parameters of the gobs 15 . In detail , it aims to reduce the variance and disparity in glass gobs 15 , due to suboptimal control and limited precision in glass manufacturing processes . In fact , this variance could lead to the emergence of defects in the produced glass containers .

[0175] The proposed solution of fers the potential to enhance product quality, reduce material waste and optimise operational ef ficiency . Furthermore , the use of metaheuristic approaches provides a flexible framework that can be adapted to various glass manufacturing settings and that could also be extended to other process control challenges beyond glass gobs control .

[0176] In detail , the present solution is based on data-driven meta-heuristic optimisers such as Genetic Algorithms , Simulated Annealing and Nelder-Mead . These meta-heuristic algorithms provide a powerful optimisation framework for searching through the vast solution space to find the optimal set of control parameters . By iteratively evaluating and updating potential solutions based on fitness criteria, these algorithms ef ficiently converge towards the optimal solution, ensuring the best possible control configuration for the given glass gob properties .

[0177] The present solution thus enhances the control o f gob properties by providing a more precise and general control schema . Additionally, it overcomes the limitations associated with current control methods for glass gobs . In fact , this solution combines an adaptable model with advanced control techniques and data-driven optimisation approaches , of fering signi ficant enhancements in the following key areas : control precision and generalisability; adaptability to process variations ; data availability and quality; and automatisation and continuous learning from history .

[0178] In particular, this solution employs data-driven models that are tailored to the glass manufacturing process , enabling them to capture the system dynamics more accurately . Unlike simplistic linear or rule-based models , these precise models are capable of accurately predicting configurations proposed by the optimisers , resulting in improved control precision and generalisability . The model utilised in this solution undergo re-training at each control step, ensuring its adaptability and precision in predicting the controlled gob properties , even with changes in parameters and environmental variables . This constant re-training and adaptation mechanism allows for ef fective control performance , regardless of process variations .

[0179] The use of the optimiser enables the system parameters to be adapted for optimal control . This adaptability introduces slight variations in the control parameters at each step, accommodating dri fts in the system and changes in the environment . Consequently, this approach mitigates the issue of data monotonicity and allows for the exploration of a wider range in the control search space , resulting in more robust and generalisable models .

[0180] Moreover, this solution introduces a model predictive control mechanism that employs machine learning and metaheuristic optimisation techniques to autonomously manage the gob parameters . By using these advanced approaches , the system dynamically adapts to varying conditions and process variables without the need for human intervention . It prevents the errors and inconsistencies brought about by manual adj ustments and individual operator expertise . The algorithms continually analyse real-time data and the optimiser adj usts the control parameters within predefined bounds to achieve the desired gob parameters with high precision . This results in a more consistent , reliable and ef ficient control system that signi ficantly reduces the need for human operators to make frequent adj ustments , thereby mitigating the ef fects of operator variability and potential errors .

[0181] Moreover, the ability to form di f ferent glass containers with gobs having di f ferent physical properties can provide numerous benefits . Such benefits include reduced mould inventory, greater production planning flexibility and longer run time (with reduced lost time for startup ) for small production runs .

[0182] Furthermore , the present solution does not perform a control based on prioritisation of the control parameters to be changed, contrary to several prior art solutions . This improves the control accuracy because the adaptability of the model and the optimiser is less constrained .

[0183] Moreover, generally the control parameters do not relate to the control timing of the components of the gob forming system 14 , that is fixed in the present solution . In other words , in the present solution the timing parameters of the gob feeder apparatus 30 and / or of the shearing mechanism 32 are not adapted based on the performed control , contrary to several prior art solutions . Therefore , these timing parameters ( such as the actuation timing of the plunger needles 50 and of the shear blades 52 ) are fixed and independent from the executed control and thus from the measured gob parameters . On the other hand, parameters such as the dead points of the plunger needles 50 and phase of the shear blades 52 can be varied based on the performed control .

[0184] Finally, it is clear that modi fications and variations may be made to what has been described and illustrated herein, without thereby departing from the scope of the present invention, as defined in the annexed claims . For example, the different embodiments described can be combined with each other to provide further solutions.

[0185] Moreover, the timing required to perform the steps of method 100 can be different with respect to what has been previously described. For example, if the re-training of the model and the determination of the optimal set of control parameters by the optimiser are performed in a time period which is shorter than a cycle time (i.e., the time period required by the gob forming system 14 for producing a set of gobs 15, starting from the receipt of the control parameters) , the re-training frequency can be unitary and thus the re-training can be executed at each sub-iteration.

Claims

CLAIMS1. A gob forming system (14) for an individual section machine (10) , the gob forming system (14) comprising:- a gob feeder apparatus (30) configured to receive a flow of molten glass (13) and to generate a plurality of molten glass streams (45) ; a shearing mechanism (32) coupled to gob feeder apparatus (30) so as to receive the molten glass streams (45) and configured to cut each molten glass streams (45) into at least one glass gob (15) ;- a gob sensing apparatus (36) operatively coupled to the gob feeder apparatus (30) and the shearing mechanism (32) and configured to measure gob parameters indicative of the glass gobs (15) ; and- a gob forming control unit (34) coupled to the gob feeder apparatus (30) , the shearing mechanism (32) and the gob sensing apparatus (36) and configured to acquire the gob parameters from the gob sensing apparatus (36) and control the gob feeder apparatus (30) and the shearing mechanism (32) based on control parameters determined based on the gob parameters , wherein the gob parameters comprise at least a gob mass, and wherein the gob forming control unit (34) is configured to implement a model and an optimiser to determine, starting from the measured gob parameters, the control parameters for controlling the gob feeder apparatus (30) and the shearing mechanism (32) , the optimiser being a meta-heuristic optimiser .

2. The gob forming system according to claim 1, whereinthe model is a non-linear model.

3. The gob forming system according to claim 1 or 2, wherein the model comprises one of the following models: a Neural Network feeder model, in particular a multi-layer perceptron model, and a Stochastic Gradient Descent model.

4. The gob forming system according to any one of the previous claims, wherein the optimiser comprises one of the following optimisers: a Nelder-Mead optimiser; a Simulated Annealing algorithm; an adaptive multi-objective simulated annealing algorithm; and a Genetic Algorithm, in particular a Non-dominated Sorting Genetic Algorithm ITT.

5. The gob forming system according to any one of the previous claims, wherein the gob feeder apparatus (30) and / or the shearing mechanism (32) are configured to be controlled based on timing parameters that are fixed and independent from the measured gob parameters.

6. The gob forming system according to any one of the previous claims, wherein the gob sensing apparatus (36) comprises a plurality of image acquisition devices (56) configured to image each one of the glass gobs (15) , the image acquisition devices (56) having respective observation directions that are inclined between them so as to image each one of the glass gobs (15) from respective points of view .

7. The gob forming system according to any one of the previous claims, wherein the gob feeder apparatus (30) comprises :- a refractory spout bowl (40) having one or a plurality of discharge orifice holes (42) located in the bottom of the spout bowl (40) , the spout bowl (40) being configured tocontain the molten glass (13) ;- a feeder tube (44) extending into the spout bowl (40) and controllable to move into the spout bowl (40) towards, and away from, the discharge orifice holes (42) so as to control the quantity of molten glass (13) into the feeder tube ( 44 ) ;- a plurality of plunger needles (50) extending within the feeder tube (44) , each one respectively above one of the discharge orifice holes (42) , the plunger needles (50) being controllable to move downwardly and upwardly in the feeder tube (44) so as to respectively advance and retract towards and away from the discharge orifice holes (42) in order to selectively advance the molten glass (13) through the discharge orifice holes (42) to generate the molten glass streams (45) , and wherein the shearing mechanism (32) comprises one or more shear blades (52) located below the discharge orifice holes (42) and controllable to cut the molten glass streams (45) into the glass gobs (15) .

8. The gob forming system according to claim 7, wherein the control parameters include at least one of the following: movement profile of the plunger needles (50) ; position of the feeder tube (44) relative to the bottom of the spout bowl (40) ; rotational speed of the feeder tube (44) relative to the bottom of the spout bowl (40) ; and phase of the one or more shear blades (52) .

9. The gob forming system according to any one of the preceding claims, wherein the gob parameters further comprise at least one of the following: gob length, measured longitudinally to the gob (15) ; gob diameter, measured in across-section of the gob (15) ; 2D gob shape; 3D gob shape.

10. An individual section machine (10) for producing glass containers (19) , comprising:- a molten glass supply (12) configured to receive batched materials (11) and to generate a flow of molten glass (13) based on the batched materials (11) ;- a gob forming system (14) according to any one of the previous claims, the gob forming system (14) being coupled to the molten glass supply (12) and being configured to receive the molten glass (13) produced by the molten glass supply ( 12 ) ;- a gob distribution arrangement (16) coupled to the gob forming system (14) and configured to receive the glass gobs (15) produced by the gob forming system (14) and distribute the glass gobs (15) to said plurality of individual sections (18a; 18b) of the individual section machine (10) ; and said plurality of individual sections (18a; 18b) coupled to the gob distribution arrangement (16) and configured each one to receive the respective one or more glass gobs (15) distributed by the gob distribution arrangement (16) and to generate a respective number of said glass containers (19) .

11. A method (100) for producing glass gobs (15) , executed by a gob forming system (14) according to any one of claims 1-9, the method (100) comprising the steps of: a. initialising (S10) , by the gob forming control unit (34) , at least the model and the optimiser; b. controlling (S12) , by the gob forming control unit (34) and for each sub-iteration of the method (100) , the gobfeeder apparatus (30) and the shearing mechanism (32) based on a respective set of said control parameters for the considered sub-iteration; c. while controlling (S12) the gob feeder apparatus (30) , acquiring (S14) , by the gob forming control unit (34) and through the gob sensing apparatus (36) , a respective set of said gob parameters in each sub-iteration, each set of gob parameters being indicative of the one or more glass gobs (15) produced by the gob forming system (14) in the respective sub-iteration; d. in an updating iteration of the method (100) , training (S16) , by the gob forming control unit (34) , the model based on a training dataset comprising the sets of the control parameters and the respective sets of the gob parameters of a plurality of said sub-iterations that precede said updating iteration; e. in said updating iteration, finding (S18) , by the gob forming control unit (34) and through the optimiser, an optimal set of the control parameters that, through the trained model, provides a respective optimal set of the gob parameters, the optimal set of the gob parameters being selected, among a plurality of candidate sets of the gob parameters examined by the optimiser, based on a comparison of the candidate sets of the gob parameters with respect to a target set of the gob parameters, the target set of the gob parameters being an input of the optimiser; and f. in said updating iteration, updating (S20) , by the gob forming control unit (34) and based on the optimal set of the control parameters, the control parameters for controlling the gob feeder apparatus (30) and the shearingmechanism ( 32 ) .12 . The method according to claim 11 , wherein the steps d- f are periodically repeated for a plurality of updating iterations , each one of said updating iterations comprising one or more respective sub-iterations , wherein the steps b and c are executed at each one of said sub-iterations , and wherein the step b is executed using the updated control parameters obtained at the precedent updating iteration .13 . The method according to claim 11 or 12 , wherein the model is initiali sed based on model initial ising parameters comprising at least one of the following : an architecture of the model ; a training acquisition number indicative of a number of sub-iterations required for collecting the sets of the control parameters and the respective sets of the gob parameters for the training dataset ; a selection of the types of gob parameters to be used, among a plurality of predefined choices ; a selection of the types of control parameters to be used, among a plurality of predefined choices ; and a model training approach, and wherein the optimiser is initialised based on optimiser initialising parameters comprising said target set of the gob parameters and at least one of the following : an architecture of the optimiser ; a selection of the types of gob parameters to be used, among a plurality of predefined choices ; a selection of the types of control parameters to be used, among a plurality of predefined choices ; an optimisation criterion of the optimiser ; and control parameter limits indicative of limits of at least some of the control parameters .14 . The method according to claim 12 or according to claims 12 and 13 , wherein the step d comprises building said training dataset according to one of the following options : ephemeral approach and cumulative approach, wherein in the ephemeral approach only the sets of the control parameters and the respective sets of the gob parameters of a subset of the most recent sub-iterations preceding the considered updating iteration are comprised in the training dataset for said updating iteration, and wherein in the cumulative approach all the sets of the control parameters and the respective sets of the gob parameters of the sub-iterations preceding the considered updating iteration are comprised in the training dataset for said updating iteration .15 . Computer program product storable in a gob forming system ( 14 ) according to anyone of claims 1- 9 , the computer program being designed so that , when executed, the gob forming system ( 14 ) becomes configured to execute a method ( 100 ) according to any one of claims 11- 14 .

Images