Method and system for inspecting smoking articles - Patents.com

Abstract

The present disclosure relates to a method for inspecting smoking articles, each smoking article including an elongated tubular wrapper (IC) extending along a longitudinal axis (AC), a filler material (MTA) wrapped in the wrapper (IC), and an elongated body (CP) embedded in the filler material (MTA) at a predetermined location. The method includes capturing an image (IMM) of the smoking article or a portion thereof observed along a light path oriented along the longitudinal axis (AC), and feeding, by a processor, the image (IMM) to a machine learning model (MA), where the machine learning model (MA) is trained to identify the body (CP) in the image (IMM).

Description

[Technical field]

[0001] The present invention relates to a method and system for inspecting smoking articles. [Background technology]

[0002] In the field of smoking articles, for example cigarettes and heat-not-burn (HNB) devices, it is necessary to carry out a quality check of the individual products. Indeed, automatic machines for producing smoking articles sometimes produce defective smoking articles. In particular, there are smoking articles that include a body arranged in a wrapper, where the body must be placed in a predetermined position. In this case, the defect may be the result of an incorrect positioning of the object. For example, some cigarettes have a filter that includes an object, the cross section of which is in the form of a curved object, as described in patent document WO2020128827A1 in the name of the applicant, said object being marked with 2 in the figure. In another example, as described in patent document WO2012016795A1, an HNB device includes an object with a circular cross section, said object being marked with 6 in the figure.

[0003] In this connection, the automatic machine for producing smoking articles may be provided with an inspection system as is known in the prior art, for example from patent document EP 3520631 A1 in the name of the Applicant.

[0004] Examples of methods for inspecting smoking articles are described in the following prior art documents: WO2018185722A1, CN111972700A, EP3476228A1, ITBO20080755A1 and EP3067823A1. Furthermore, with regard to assembly machines for manufacturing smoking articles, the prior art provides solutions, for example, as described in documents DE102014203158A1 and EP0653170A1.

[0005] However, there remains a need for a system which is particularly efficient, reliable and rapid in carrying out inspection of smoking articles, and in particular it would be desirable for such a system to be able to carry out reliable quality inspection in-process during manufacture. Summary of the Invention

[0006] SUMMARY OF THE PRESENT EMBODIMENT It is an object of the present invention to provide a method and system for inspecting smoking articles which overcomes the above-mentioned drawbacks of the prior art.

[0007] This object is fully achieved by the method and system according to the invention as characterized in the appended claims.

[0008] According to one aspect, the present disclosure provides a method for inspecting smoking articles, each smoking article including an elongated tubular wrapper extending along a longitudinal axis, a filler material wrapped in the wrapper, and an elongated body embedded in the filler material at a predetermined location.

[0009] The method includes capturing an image of the smoking article, or a portion thereof, which is viewed along a light path oriented along the longitudinal axis.

[0010] The method includes providing an image to a machine learning model, the machine learning model being trained to identify bodies in the image, the step being performed by a processor.

[0011] This allows the machine to check (at least) the presence of the body within the filling material.

[0012] Preferably, the machine learning model is trained to generate positioning data for each image, the positioning data representing the position of the body relative to the tubular wrapper. The positioning data represents a planar geometric figure that defines the boundary of the object represented in the image. In this way, it is possible to not only check the presence of the body in the filling material, but also to identify the position of the body relative to the wrapper.

[0013] The method includes processing the positioning data by a processor based on reference data representative of a predetermined position of the body, e.g., the processing includes a step performed by the processor of comparing the positioning data with the reference data representative of a predetermined position of the body.

[0014] Thus, the processor may check that the body is effectively positioned in an acceptable position relative to the predetermined position. For example, the processor may have access to a deviation value identifying a maximum deviation between a position of the body determined from the positioning data and the predetermined position.

[0015] In one embodiment, the machine learning model comprises a deep neural network, hi one embodiment, the neural network is a convolutional neural network.

[0016] The neural network may include one or more filtering stages of max-pooling type.

[0017] The machine learning model is trained by supplying a plurality of exemplar images, and for each image, the machine learning model is trained by supplying positioning target data representing planar geometric figures that bound the objects represented in the images.

[0018] The exemplars include a first plurality of images of semi-finished smoking articles with the objects correctly positioned, and in one embodiment, the exemplars include a second plurality of images of semi-finished smoking articles with the objects incorrectly positioned in accordance with a selection of predetermined positioning defects.

[0019] The method comprises steps executed by a control unit to determine a positive or negative identification result identifying the object in the image. The identifying step is executed by a machine learning model. In case of a negative identification result, the method comprises steps executed by the control unit to generate information for rejecting the smoking article. In case of a positive identification result, the method comprises initiating a step of comparing the positioning data with reference data representing a predefined position of the body. The method then comprises generating information for rejecting or accepting the smoking article depending on the comparison. This makes it possible, firstly, to reject articles without a body and, secondly, if a body is present, to reject articles with an incorrectly positioned body.

[0020] In one embodiment, the method includes optically inspecting the wrapper for each captured image. The method includes generating wrapper data representative of the inspection of the wrapper.

[0021] The method includes the step of comparing the wrapper data with reference wrapper data representative of predetermined specifications for the wrapper, and the method includes the step of proceeding to the step of (i) generating or (ii) providing information to reject in response to (i) a negative result or (ii) a positive result, respectively, of the comparing step.

[0022] In this way, the method involves a complete inspection of the smoking article, including checking the wrapper.

[0023] According to one aspect, the method is for inspecting smoking articles one by one. In other words, the method comprises capturing an image of each individual smoking article. Preferably, the method comprises the step of placing an individual smoking article in an inspection zone, the inspection device being targeted at the smoking article in order to obtain an image of the smoking article. In one embodiment of the method, the inspection zone is illuminated with constant illumination parameters to allow uniform illumination of the smoking article, and the illumination parameters of the inspection zone are the same as those used to capture the plurality of exemplar images. In one embodiment of the method, the smoking articles are placed in the inspection zone with the same orientation relative to the inspection device, and the smoking articles depicted in the plurality of exemplar images have the same orientation as the smoking articles inspected during the manufacturing process.

[0024] Uniformity of lighting and positioning between the captured images and the exemplar images greatly improves the reliability of inspection by the machine learning model, which would not be possible if inspections were performed in groups rather than individually, with each smoking article having its own unique position relative to the inspection device.

[0025] In one embodiment, the inspection device is synchronized with a conveyor that transports the smoking articles (or semi-finished products or parts used to construct the smoking articles) into and out of the inspection zone. Preferably, the smoking articles (or semi-finished products or parts used to construct the smoking articles) are transported one by one in an orderly manner. Preferably, they are transported to the inspection station such that their orientation is predetermined and known to the system. This improves the inspection accuracy and facilitates the training of machine learning models for defect recognition or quality control. Furthermore, for each inspected smoking article, the control unit knows the result of the inspection. As a result, the control unit can control the rejection device to reject single defective smoking articles.

[0026] It should be noted that individualizing smoking articles also facilitates the rejection effort: indeed, in a solution in which smoking articles are inspected in groups, not only is it more difficult to identify a single defective smoking item, but it is also more complicated to reject it, because (i) smoking articles need to be retroactively individualized, or (ii) multiple smoking items need to be rejected because of the presence of a single defective smoking item.

[0027] In at least one exemplary embodiment, the smoking articles or parts thereof are moved individually as a single article along a predetermined path in an apparatus for manufacturing smoking articles (in one example an assembly apparatus, but for purposes of this and other aspects of the present disclosure relating to the inspection system and image processing mode, it could be a different type of apparatus). At least one inspection stage (e.g. a first inspection stage, or a first and a second inspection stage) is arranged along the predetermined path. The inspection stage includes at least one illuminator and one camera. Preferably, as also explained above, the smoking articles or parts thereof are moved along the predetermined path one by one. Preferably, the smoking articles or parts thereof are also moved along the predetermined path such that they have a predetermined orientation in the inspection stage (the predetermined orientation is known to the system with respect to the optical path of the camera, e.g. the smoking articles or parts thereof are oriented to be aligned with the camera).

[0028] In one example, during the step of capturing images, each smoking article or part thereof moved along the predetermined path is individually viewed and photographed (i.e. subject to image capture), such that each image relates to a single smoking article or part thereof.

[0029] It should be noted that with respect to the machine learning model, it is preferably trained by supplying it with a number of exemplar images, each relating to a single smoking article or part thereof. Furthermore, the machine learning model may be trained by supplying it with target data (in which case the learning is supervised). Alternatively, no target data is supplied to it (in which case the learning is unsupervised).

[0030] In one example, the machine readable instructions contained in the non-transitory data storage are such that for each smoking article or portion thereof in the flow, the processor: - capturing a plurality of images of the smoking article, or a part thereof, observed along a path having a predetermined orientation through at least one inspection stage (e.g. through a first and a second inspection stage); -Feed multiple images into a machine learning model.

[0031] The machine learning model is trained to identify a predefined category of defect of the smoking article or part thereof. In one example, the category of defect relates to the location or lack thereof of a body located (at a predefined location) within the smoking article. However, the machine learning model may be trained to recognise other types of defects, such as, by way of example only, the presence of dirt or impurities, or a shape that differs from a predefined (reference) shape of the smoking article or part thereof.

[0032] According to one aspect, the present disclosure provides a computer program comprising instructions for performing the steps of the methods described in the present disclosure when executed by a processor.

[0033] According to one aspect, the present disclosure provides an inspection system for inspecting smoking articles including an elongated tubular wrapper extending along a longitudinal axis, a filler material enclosed in the wrapper, and an elongated body embedded in the filler material at a predetermined location. The system comprises a processor and non-transitory data storage including machine-readable instructions. The machine-readable instructions direct the processor to capture an image of the smoking article or a portion thereof observed along a light path oriented along the longitudinal axis. The machine-readable instructions direct the processor to feed the image to a machine learning model that is machine-trained to identify the body in the image.

[0034] The non-transitory data storage includes further instructions that direct the processor to process the images to generate, for each image, positioning data representative of a planar geometric figure that defines the boundary of the body represented in that image.

[0035] The non-transitory data storage includes further instructions that direct the processor to compare the positioning data to reference data representative of a predetermined position of the body.

[0036] The instructions direct the processor to determine whether an identification of an object in an image performed by a machine learning model is positive or negative. If the identification is negative, the instructions direct the processor to generate information for rejecting the smoking article. If the identification is positive, the instructions direct the processor to compare the positioning data to reference data representative of a predetermined position of the body and generate information for rejecting or accepting the smoking article in response to the comparison.

[0037] In one embodiment, the machine learning model includes a deep neural network. The deep neural network may be a convolutional deep neural network. The deep neural network may include one or more filtering stages of max pooling type.

[0038] According to one aspect, the present disclosure provides an apparatus for continuous cycle production of smoking articles, each smoking article including an elongated tubular wrapper extending along a longitudinal axis, a filler material enclosed in the wrapper, and an elongated body embedded in the filler material at a predetermined location, the apparatus comprising an inspection apparatus according to any of the features described in the present disclosure.

[0039] According to one aspect, the present disclosure provides a method for inspecting smoking articles, each smoking article including a body located at a predetermined position within the smoking article, the method being applicable in the context of operations performed by an apparatus for manufacturing smoking articles in a continuous cycle through a series of processing stages from an initial processing stage to a final processing stage.

[0040] The method includes capturing an image of a workpiece produced by a machine at an intermediate processing stage between an initial processing stage and a final processing stage.

[0041] The method includes steps performed by a control unit to provide images to a machine learning model that is trained to identify bodies in the images.

[0042] According to one aspect, the present disclosure provides an assembly apparatus for producing multi-component smoking articles, each of the smoking articles including a plurality of rod sections defining a respective central axis, the plurality of rod sections including a first rod section with a flavor element and a second rod section.

[0043] The assembly apparatus includes a coupling unit configured to form rod section groups, each group including at least a first rod section and a second rod section, the first and second rod sections being axially aligned with one another, and the first and second rod sections being in end-to-end abutment.

[0044] The rod sections advance perpendicular to their central axes, and the first rod section is derived from a rod each having a first end and a second end spaced apart along the rod axis.

[0045] The coupling unit is configured to couple respective second rod sections to the first and second ends of the rod. The coupling unit is configured to separate the rod into a pair of parts by cutting the rod transversely to the rod axis to form a corresponding pair of rod section groups in a separation stage. Each rod section group includes a part of the rod and a second rod section.

[0046] The assembly apparatus includes a packaging unit configured to receive the series of rod sections fed in a feed direction from the combining unit, feed the rod sections perpendicular to their central axes, and wrap a sheet of wrapping material around each rod section.

[0047] The assembly apparatus includes an inspection system.

[0048] The inspection system includes a first inspection stage disposed upstream of the separation stage in the feed direction, the first inspection stage configured to axially view the first and second ends of each rod.

[0049] The inspection system includes a second inspection stage disposed downstream of the separation stage in the feed direction, the second inspection stage configured to axially view a free end of a first rod section of each rod section of the pair of rod section groups.

[0050] In one embodiment, the assembly apparatus includes a spacing stage disposed between the separation stage and the second inspection stage in the feed direction, the spacing stage configured to axially space the pair of rod sections.

[0051] The assembly apparatus is configured to feed a pair of rod sections perpendicular to the rod axis between the separation stage and the spaced apart stage.

[0052] The second inspection stage includes a central camera axially disposed between the pair of rod sections. In one embodiment, the second inspection stage includes a further central camera axially disposed between the pair of rod sections.

[0053] In one embodiment, the second inspection stage includes a right-side inspection unit including a central camera facing to the right relative to the feed direction for viewing the free end of one of the pair of rod sections.

[0054] In one embodiment, the second inspection stage includes a left-hand inspection unit including a further central camera facing leftward relative to the feed direction for observing the free end of the other rod section group of the pair of rod sections.

[0055] In a preferred embodiment, the right and left inspection units are offset along the feed direction.

[0056] Preferably, the first inspection stage is provided with a pair of cameras.

[0057] In one embodiment, the pair of cameras have respective optical paths oriented in opposite convergence directions. Additionally, according to one aspect of the present disclosure, a second pair of cameras have respective optical paths oriented in aligned but opposite convergence directions.

[0058] The assembly apparatus comprises a control unit. The control unit is connected to and drives the first inspection stage and the second inspection stage. The control unit is provided with a processor and non-transitory data storage including machine readable instructions that direct the processor to capture a plurality of images of the smoking article, or a portion thereof, viewed along a light path oriented along the longitudinal axis through the first and second inspection stages. It should be noted that the smoking article includes a body disposed at a predetermined location within the smoking article.

[0059] The control unit is provided with a processor and non-transitory data storage including machine-readable instructions that direct the processor to feed a plurality of images to a machine learning model that is trained to identify bodies in the images.

[0060] According to one aspect, the present disclosure provides an apparatus for manufacturing smoking articles, preferably multi-component articles, each smoking article comprising parts that are inspectable by optical equipment in at least one inspection station of the apparatus and that are hidden from the optical equipment once the smoking article is completed. In other words, the present disclosure relates to an apparatus in which the smoking articles are obtained by a progressive process that defines inspectable semi-finished products that cannot be inspected by optical equipment once the smoking article is completed.

[0061] These machines are equipped with inspection systems, which are arranged in inspection stations of the machines corresponding to the semi-finished products that can be inspected.

[0062] The apparatus comprises a first processing station configured to produce an inspectable semi-finished product from the intermediate part, and a second processing station configured to produce smoking articles from the inspectable semi-finished product.

[0063] The inspection station is located along a supply path of the smoking articles through the apparatus between the first and second processing stations. Advantageously, the apparatus comprises a control unit. The control unit is connected to and drives the inspection system. The control unit is provided with a processor and non-transitory data storage containing machine readable instructions.

[0064] The control unit is programmed to receive a plurality of images of an inspectable workpiece or part thereof captured by the inspection system, the images being captured, for example, along a light path preferably oriented along the longitudinal axis of the smoking article, and by way of example only, the inspection system captures an image of the body at a predetermined position within the smoking article.

[0065] However, the inspection system may be configured to identify other aspects of the smoking article that are not visible once completed, for example the presence of dust or stains on the surface before it is wrapped, or the length of a rod section before it is glued to another rod section.

[0066] The control unit is programmed to feed the plurality of images to a machine learning model.

[0067] For example, but not necessarily, a machine learning model may be trained to identify the body within a smoking article in an image. In other embodiments, it may identify, for example, the presence of dust or stains on the surface before it is wrapped, or the length of a rod section before it is glued to another rod section.

[0068] According to one aspect, the present disclosure provides an assembly method for producing a multi-component smoking article, each smoking article comprising a plurality of rod sections defining a respective central axis, the plurality of rod sections including a first rod section comprising a flavor element and a second rod section.

[0069] The method includes steps performed by a combining unit to supply a stream of rods in a supply direction transverse to a rod axis, each rod having a flavor element and elongated along the rod axis from a first end to a second end.

[0070] The method includes coupling each second rod section to a first end and a second end of the rod.

[0071] The method includes a step performed by a combining unit at a separation stage of separating the rod into a pair of parts by cutting the rod transversely to the rod axis to form a corresponding pair of rod section groups.

[0072] Each rod section group includes a rod portion defining a first rod section and a second rod section axially aligned along a central axis and abutted end-to-end.

[0073] The method includes the step of feeding the rod sections perpendicular to their central axes using a packaging unit.

[0074] The method includes the step of providing rod sections perpendicular to their central axes.

[0075] The method includes wrapping a sheet of wrapping material around the group of rod sections using a wrapping system.

[0076] The method includes a first step of observing first and second ends of each rod at a first inspection stage located upstream of the separation stage in the feed direction.

[0077] The method includes a second step of observing, in a second inspection stage located downstream of the separation stage in the feed direction, a free end of a first rod section of each rod section group of a pair of rod section groups.

[0078] In one embodiment, the method further comprises the step of axially separating the pair of rod sections after the separating step and before the second observing step.

[0079] According to one aspect of the present disclosure, during the spacing step, the pair of rod sections are also fed transversely to their axes.

[0080] In one embodiment of the method, the binding step precedes the separating step.

[0081] In one embodiment, the second step of observing comprises the following sub-steps in sequence: - observing, by a central camera, a free end of one of the pair of rod sections; - feeding the rod sections in a feed direction; - observing, by a further central camera, the free end of the other of the pair of rod section groups.

[0082] In one embodiment, the smoking article includes a body located at a predetermined position within the smoking article, and the method further comprises feeding the images captured at the first and second inspection stages to a machine learning model that is trained to identify the body in each image. [Brief description of the drawings]

[0083] These and other characteristics will become more apparent from the following description of preferred embodiments, illustrated by way of non-limiting example in the accompanying drawings, in which: [Figure 1] 1 shows a schematic side view of an apparatus for the continuous cycle production of smoking articles. [Diagram 2] 2 shows a schematic top view of the device of FIG. 1. [Diagram 3] 1 illustrates diagrammatically a method for inspecting smoking articles; [Figure 4] 1 shows a schematic cross-section of a smoking article including a filler material and a body CP. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0084] With reference to the accompanying drawings, reference numeral 1 indicates an assembly apparatus for producing multi-component smoking articles. Each smoking article comprises a plurality of rod sections SP defining a respective central axis. The plurality of rod sections includes a first rod section SP1 provided with a flavour element and a second rod section SP2. The smoking article also includes an outer wrapper IC.

[0085] The apparatus 1 comprises a coupling unit 10 configured to form rod section groups GS, each including at least a first rod section SP1 and a second rod section SP2, which are axially aligned and abutted end-to-end, in which the rod section groups GS advance perpendicularly to their central axis AC.

[0086] In particular, the apparatus 1 comprises a dividing unit 11 configured to form a plurality of first rod sections SP1 and second rod sections SP2 from corresponding rods, the plurality of first rod sections SP1 and second rod sections SP2 being subsequently grouped into pairs to define rod section groups.

[0087] In particular, in one embodiment, the second rod section SP2 comprises a first portion SP2' and a second portion SP2''.

[0088] The dividing unit 11 comprises a first separating element 111 configured to separate the first rod into a corresponding rod B1 defining a first portion SP2' of the second segment SP2 and having a predetermined length along its central axis AC.

[0089] The dividing unit 11 comprises a second separating element 112 configured to separate the second rod into a corresponding rod B2 defining a second portion SP2'' of the second rod section SP2 and having a predetermined length along its central axis AC.

[0090] The dividing unit 11 comprises a third separating element 113 configured to separate the third rod into corresponding rods B3 defining a first rod section SP1 and having a length along its central axis AC. It should be noted that the third rod B3 is a rod including a flavor filler MTA. Furthermore, the first rod section SP1 should ideally include a body CP embedded in the flavor filler.

[0091] The apparatus 1 comprises an inspection system 12 arranged to capture inspection data which is processed to derive information about the smoking article.

[0092] The inspection system includes a first inspection stage 121 configured to capture inspection data regarding rods B3 defining a first rod section SP1. In particular, the rods B3 each include a first end B31 and a second end B32. The first inspection stage 121 includes a first camera pair 122 each configured to capture corresponding images of the first end B31 and the second end B32 of the rod B3.

[0093] In a first inspection stage, a first camera pair 122 captures image data 122' representative of an image of a first end B31 of rod B3 and an image of a second end B32 of rod B3.

[0094] The apparatus 1 comprises a control unit 2 , which includes a processor and is configured to control the apparatus 1 and to process data received from an inspection system 12 .

[0095] The control unit 2 is configured to recognize the presence or absence of the body CP in the rod B3. Furthermore, in one embodiment, the control unit 2 is also configured to check that the body CP is in the correct position relative to the filler material MTA of the rod B3.

[0096] The step of recognizing the body CP is performed by feeding the image data captured by the first inspection stage 121 through the first camera pair 122 to a machine learning model MA. The machine learning model MA is trained to identify the body CP in the images. The recognizing step is explained in more detail below.

[0097] It should then be noted that in this step the apparatus is provided with a first plurality of rods B1 and a second plurality of rods B2 as well as a third plurality of rods B3, the first plurality of rods B1 and the second plurality of rods B2 together defining a second rod section SP2, and the third plurality of rods B3 defining a first rod section SP1.

[0098] At this point, the combining unit 10 starts to operate to combine the three rods B1, B2, B3 together. The combining unit is configured to feed the rods B1, B2, B3 along a feed direction DA perpendicular to the central axis AC of the rods B1, B2, B3.

[0099] The joining unit comprises a first stage 101. The first stage 101 comprises a first cutting station 101T configured to cut the rod B1 along a direction perpendicular to the central axis AC, i.e. along the feed direction DA, to form a first segment B1' and a second segment B1'', which are separated by a first separator 101S configured to separate the first and second segments B1', B1'', along a separation direction DD perpendicular to the feed direction DA and parallel to the central axis of the (already separated) rod B1. Separating the two segments is necessary to be able to insert a rod B2 of the second plurality of rods B2 at an intermediate position between the first segment B1' and the second segment B1'', along the separation direction DD.

[0100] Next, the joining unit 10 includes a second stage 102 in which a second rod B2 of the second plurality of rods B2 is positioned at an intermediate position between the first segment B1' and the second segment B1'' along the separation direction DD.

[0101] The second stage 102 includes a first bonding device 102I configured to connect a first end of rod B2 to a corresponding end of a first segment B1' of rod B1 and to connect a second end of rod B2 to a corresponding end of a second segment B1'' of rod B1.

[0102] The second stage comprises a second cutting element 102T configured to separate the second rod B2 into a corresponding first segment B2' connected to a first segment B1' of the first rod B1 to define a first rod section SP1 of the smoking article and a second segment B2'' connected to a second segment B1'' of the first rod B1 to define a further second rod section SP2 of a further smoking article.

[0103] The second stage 102 comprises a second separator 102S configured to separate the second rod section SP2 from a further second rod section SP2.

[0104] The joining unit 10 comprises a third stage 103 in which one third rod B3 of the third plurality of rods B3 is arranged at an intermediate position along the separation direction DD between the second rod section (formed from the first segment B1' of the first rod B1 and the first segment B2' of the second rod B'') and the second further rod section (formed from the second segment B1'' of the first rod B1 and the second segment B2'').

[0105] The third stage 103 comprises a second gluing device 103I configured to connect a first end of the rod B3 to a corresponding end of the second rod section SP2, i.e. to the end of the first segment B2' of the second rod B2. The second gluing device 103I is configured to connect a second end of the rod B3 to a corresponding end of the further second rod section SP2, i.e. to the end of the second segment B2'' of the second rod B2.

[0106] The third stage 103 comprises a third cutting element 103T configured to separate the third rod B3 into a corresponding first segment B3' connected to the first segment B2' of the second rod B2 to define a first rod section group GS1 (formed from a first rod section SP1 and a second rod section SP2 defined by a first segment B3' of the third rod B3) and a second segment B3'' connected to the second segment B2'' of the second rod B2 to define a second rod section group GS2 (formed from a first rod section SP1 and a further second rod section SP2 defined by a second segment B3'' of the third rod B3).

[0107] The third stage 103 includes a third separator 103S configured to separate the first rod section group GS1 and the second rod section group GS2 along the separation direction DD.

[0108] Thus, after the third stage, there are two groups of rod sections in the joining unit 10 that are spaced apart from each other along the spacing direction DD and that are subsequently wrapped to form two smoking articles. At this stage, both ends of the first group of rod sections GS1 and the second group of rod sections GS2 have not been checked for the presence of a body CP in the filler material MTA.

[0109] Therefore, the device 1 expects a further check to check for the presence of the body CP.

[0110] Thus, the apparatus 1 includes a second inspection stage 13 configured to capture image data representing an image of the outer end of the first rod section SP1 of the first rod section group GS1 and the outer end of the first rod section SP1 of the second rod section group GS2.

[0111] In particular, the second inspection stage 13 is configured to check that the body CP extends to the outer end of the first rod section SP1 of the first rod section group GS1 and to the outer end of the first rod section SP1 of the second rod section group GS2.

[0112] In one embodiment, the second inspection stage 13 comprises a second camera pair 131 including a right camera 131A and a left camera 131B. In one embodiment, the right camera 131A is positioned spaced apart from the left camera 131B along the feed direction DA. Furthermore, the cameras 131A, 131B of the second camera pair 131 are each configured to capture image data of an outer end of the first rod section SP1 of the first rod section group GS1 and an outer end of the first rod section SP1 of the second rod section group GS2, respectively.

[0113] In another embodiment, the two cameras of the camera pair 131 are co-located along the feed direction DA and point in opposite directions.

[0114] It should therefore be noted that in a preferred embodiment, the cameras of the first camera pair are each oriented to capture images of two opposing ends of the third rod B3, while the cameras of the second camera pair 131 are oriented in the opposite direction to capture images of the free ends of the first rod section SP1 of the first and second rod section groups GS1, GS2.

[0115] In one embodiment, the second camera pair 131 is positioned above the conveyor and faces in a direction directed outwards from the conveyor. In other words, the second camera pair 131 is positioned between the first rod section group GS1 and the second rod section group GS2 along the distance direction DD. This allows for frame zones of smoking articles facing each other and therefore not visible from one side of the conveyor of smoking articles.

[0116] The apparatus 1 comprises a packaging device 14 configured to package the first group of rod sections GS1 and the second group of rod sections GS2. At the end of the second inspection stage 13, the first and second groups of rod sections GS1, GS2 are processed by the packaging device 14 which packages each of them to define a smoking article.

[0117] In one embodiment, the apparatus comprises a third inspection stage 15 configured to control the packaging of the smoking articles. In particular, the third inspection stage 15 comprises a camera configured to capture images of the wrapper IC of the smoking article. These images are transmitted to the control unit 2 which is configured to process them and compare them with reference images to check that the wrapper IC complies with predefined quality standards.

[0118] In one embodiment, the apparatus 1 comprises an unwrapping device 18 configured to unwrap smoking articles or rods B3 that do not meet prescribed quality requirements.

[0119] In particular, the control unit 2 is connected to the unwrapping device 18 for transmitting a drive signal representative of information regarding whether or not to unwrap the smoking article or rod B3. The drive signal is generated by the control unit based on information received from the first inspection stage 121 and / or the second inspection stage 13 and / or the third inspection stage 15.

[0120] Described in detail below is the inspection method implemented by the device 1 and combination unit of the present disclosure.

[0121] The method comprises capturing an image IMM of a smoking article or part thereof, namely the rod B3, the image being observed along an optical path oriented along a longitudinal axis L parallel to and coincident with the central axis AC.

[0122] The method comprises a step of feeding the image IMM to a machine learning model MA, which is trained to identify the body CP in the image. This step is performed by a processor of the control unit 2.

[0123] The machine learning model MA is trained to generate, for each image, positioning data 201. The positioning data 201 represent the position of the body CP relative to the tubular wrapper. The positioning data 201 represent a planar geometric figure that defines the boundary of the object represented in the image IMM. In this way, in addition to checking the presence of the body CP in the filling material MTA, it is possible to identify the position of the body CP relative to the wrapper.

[0124] The method comprises a processor-executed step FCN of comparing the positioning data 201 with reference data 201' representative of a predefined position of the body CP. The machine learning model MA comprises the reference data 201' which determines its processing in the identifying step (i.e. the machine learning model MA prescribes an identification algorithm based on the reference data 201'). The reference data 201' is derived automatically by the machine learning model MA in the learning step from a number of training exemplars (comprising corresponding positioning data vectors associated with respective target values). Preferably, the training exemplars comprise a number of exemplar images and corresponding target values.

[0125] In one embodiment, the machine learning model MA comprises a deep neural network, hi one embodiment, the neural network is a convolutional neural network.

[0126] The neural network may include one or more filtering stages of max-pooling type.

[0127] The machine learning model MA is trained by supplying it with a number of exemplar images, and for each image the machine learning model is trained by supplying it with positioning target data representing a planar geometric figure that bounds the object represented in the image.

[0128] The exemplar comprises a first plurality of images IM1 of a semi-finished smoking article in which the object is correctly positioned. In one embodiment, the exemplar comprises a second plurality of images IM2 of a semi-finished smoking article in which the object is incorrectly positioned in accordance with a selection of predetermined positioning defects.

[0129] The method comprises a step FDT, executed by the control unit 2, of determining a positive or negative identification result for identifying the object in the image. The identification is performed by a machine learning model MA. In case of a negative identification result 202', the method comprises a step FGN, executed by the control unit 2, of generating information 203 for rejecting the smoking article. In case of a positive identification result 202'' the method comprises a step FCN, executed by the control unit 2, of initiating a step of comparing the positioning data with reference data representative of a predefined position of the body CP. Then the method comprises generating information 203 for rejecting or accepting the smoking article depending on the comparing step FCN. [Prior art documents] [Patent documents]

[0130] [Patent Document 1] WO2020128827A1 [Patent Document 2] WO2012016795A1 [Patent Document 3] EP3520631A1 [Patent Document 4] WO2018185722A1 [Patent Document 5] CN111972700A [Patent Document 6] EP3476228A1

Patent document 7

Patent document 8

Patent document 9

Patent document 10

Claims

1. 1. A method for inspecting smoking articles, each smoking article comprising an elongated tubular wrapper (IC) extending along a longitudinal axis (AC), a filler material (MTA) enclosed in said wrapper (IC), and an elongated body (CP) embedded in said filler material (MTA) at a predetermined location, the method comprising: - capturing an image (IMM) of said smoking article or part thereof as viewed along an optical path oriented along said longitudinal axis (AC); - feeding, by a processor, the image (IMM) to a machine learning model (MA), wherein the machine learning model (MA) is trained to identify the body (CP) in the image (IMM).

2. 2. The method of claim 1, wherein the machine learning model (MA) is trained to generate, for each image (IMM), positioning data (201) representing planar geometric figures that bound the objects represented in that image (IMM).

3. 3. A method according to claim 2, comprising a step (FCN) of processing, by said processor, said positioning data (201) on the basis of reference data (201') representative of said predetermined position of said body (CP).

4. The method of claim 1 , wherein the machine learning model (MA) comprises a deep neural network.

5. The method of claim 4 , wherein the neural network is a convolutional neural network.

6. The method of claim 4 , wherein the neural network includes one or more filtering stages of max-pooling type.

7. The machine learning model (MA) - providing a number of example images (IM1, IM2); - for each image, providing positioning target data representing a planar geometric figure that bounds the object represented in that image.

8. The method of claim 7, wherein the plurality of exemplary images includes a first plurality of images (IM1) of semi-finished smoking articles in which the objects are correctly positioned and a second plurality of images (IM2) of semi-finished smoking articles in which the objects are incorrectly positioned in accordance with a predetermined selection of positioning defects.

9. - a step (FDT) of determining, by a control unit (2), a positive identification result (202'') or a negative identification result (202') of said body (CP) in said image (IMM), said step of identifying being performed by said machine learning model (MA); - generating, by said control unit (2), in case of a negative identification result (202'), information (203) for rejecting said smoking article; A method according to any one of claims 1 to 3, comprising a step (FGN) of comparing (FCN) by the control unit (2) in the case of a positive identification result (202'') the positioning data (201) with the reference data (201') representing the predetermined position of the body, and generating information (FGN) for rejecting (203) or approving the smoking article depending on the step (FCN) of comparing.

10. For each image captured, - optically inspecting said wrapper (IC) and generating wrapper data representative of the inspection of said wrapper; - comparing said wrapper data with reference wrapper data representing predetermined specifications for said wrapper; The method of claim 9, further comprising the step of: - depending on (i) a negative result or (ii) a positive result of said comparing step, respectively: (i) generating information for excluding (203) or (ii) proceeding to a step of feeding said image (IMM) to a machine learning model (MA).

11. 4. A method according to any one of claims 1 to 3, wherein during the capturing step, the smoking articles or parts thereof are moved individually along a predetermined path, whereby each of the smoking articles or parts thereof moved along the predetermined path is individually observed and subject to image capture.

12. The machine learning model (MA) - providing a number of example images (IM1, IM2); - for each image, providing target data; 4. A method according to any one of claims 1 to 3, wherein each example image relates to a single smoking article or part thereof.

13. A computer program comprising instructions for carrying out the steps of the method according to any one of claims 1 to 3 when executed by the processor.

14. 1. An inspection system (12) for inspecting smoking articles comprising an elongated tubular wrapper (IC) extending along a longitudinal axis (AC), a filler material (MTA) wrapped in the wrapper (IC), and an elongated body (CP) embedded in the filler material (MTA) at a predetermined location, the inspection system comprising a processor and non-transitory data storage including machine-readable instructions that instruct the processor to: - instructing the capturing of an image (IMM) of said smoking article or part thereof as viewed along an optical path oriented along said longitudinal axis (AC); - an inspection system (12) that instructs the image (IMM) to be provided to a machine learning model (MA), the machine learning model (MA) being trained to identify the body (CP) within the image (IMM).

15. 15. The inspection system of claim 14, wherein the non-transitory data storage includes further instructions that instruct the processor to process the images to generate, for each image, positioning data (201) that represent a planar geometry that bounds the body (CP) represented in that image (IMM).

16. 16. The inspection system of claim 15, wherein the non-transitory data storage includes further instructions directing the processor to compare the positioning data (201) with reference data (201') representative of the predetermined position of the body (CP).

17. The non-transitory data storage includes further instructions configured to direct the processor to: - determining a positive or negative identification result (202'', 202') identifying the object in said image (IMM), said step of identification being performed by a machine learning model (MA); - if the identification result is negative (202'), command the generation of information (203) for rejecting said smoking article; - An inspection system as claimed in any one of claims 14 to 16, which, if the identification result is positive (202''), instructs to compare the positioning data (201) with the reference data (201') representing the predetermined position of the body (CP) and to generate information (203) for rejecting or approving the smoking article depending on the comparing step (FCN).

18. 17. The inspection system of claim 14, wherein the machine learning model (MA) comprises a deep neural network, the deep neural network being a convolutional network and including one or more filtering stages of max-pooling type.

19. 17. An inspection system as claimed in any one of claims 14 to 16, wherein the smoking articles or parts thereof are moved individually along a predetermined path by a conveyor generating a stream of smoking articles or parts thereof moving along the predetermined path, and the processor is programmed to capture the image (IMM) of a single smoking article or part thereof observed individually for each smoking article or part thereof in the stream.

20. An apparatus for the continuous cycle production of smoking articles, each smoking article comprising an elongated tubular wrapper (IC) extending along a longitudinal axis (AC), a filler material (MTA) wrapped in the wrapper, and an elongated body (CP) embedded in the filler material (MTA) at a predetermined position, the apparatus comprising an inspection system (12) as described in any one of claims 14 to 16.

21. 1. A method for inspecting smoking articles, each smoking article comprising a body (CP) located at a predetermined position within the smoking article, the method comprising: - in an apparatus for continuously cyclically producing said smoking articles through a series of processing stages from an initial processing stage to a final processing stage, capturing machine-generated images (IMM) of semi-finished products at an intermediate processing stage between said initial processing stage and said final processing stage; - feeding, by a control unit (2), said image (IMM) to a machine learning model (MA), wherein said machine learning model (MA) is trained to identify said body (CP) in said image (IMM).

22. 1. An assembly apparatus (1) for producing multi-component smoking articles, each smoking article comprising a plurality of rod sections defining a respective central axis (AC), said plurality of rod sections comprising a first rod section (SP1) and a second rod section (SP2) comprising a flavour element, said assembly apparatus (1) comprising: a joining unit (10) configured to form groups of segments (GS1, GS2), each of said groups of segments (GS1, GS2) comprising at least said first segment (SP1) and said second segment (SP2) aligned axially and abutting end to end, said groups of segments (GS1, GS2) being fed perpendicular to their central axis (AC), said first segments (SP1) being obtained from bars (B3) each having first and second ends (B31, B32) spaced apart along the bar axis; a coupling unit (10) configured to couple the first and second ends (B31, B32) of the bar to a corresponding second segment (SP2), and during a separation stage (103T), separate the bar into a pair of portions (B3', B3'') by cutting the bar transversely to the bar axis (AC) to form a corresponding pair of segment groups (GS1, GS2), each segment group (GS1, GS2) comprising a bar portion (B3', B3'') and a second segment (SP2); a wrapping unit (14) configured to receive from said combining unit (10) a succession of segments fed in a feed direction, to feed said segments perpendicular to their central axis (AC) and to wrap a sheet of wrapping material around each segment; An inspection system (12) according to any one of claims 14 to 16, a first inspection stage (121) arranged upstream of the separation stage (103T) in the feeding direction, the first inspection stage (121) comprising a first pair of cameras (122) adapted to axially observe the first and second ends (B31, B32) of each bar (B3); - an inspection system (12) including a second inspection stage (13) arranged downstream of the separation stage (103T) in the supply direction, the second inspection stage (13) having a second camera pair (131) and configured to axially observe the free end of the first segment (SP1) of each segment group of the pair of segment groups (GS1, GS2).

23. 23. The assembly apparatus (1) of claim 22, further comprising a spacing stage (103S) arranged between the separation stage (103T) and the second inspection stage (13) in the supply direction and configured to axially space the pair of rod section groups (GS1, GS2).

24. 24. The assembly device (1) according to claim 23, wherein the assembly device (1) is configured to supply the pair of segment groups (GS1, GS2) between the separation stage (103T) and the spacing stage (103S) in a direction perpendicular to the bar axis (AC).

25. 23. The assembly apparatus (1) of claim 22, wherein the cameras of the first camera pair (122) have respective optical paths oriented in opposite converging directions, and the cameras of the second camera pair (131) have respective optical paths oriented in opposite diverging directions.

Images