Packaging line computer-aided troubleshooting using large language models

Abstract

A method (600) for computer-aided troubleshooting of a packaging line (1) configured to produce packages is disclosed. In one embodiment, the method (600) comprises the steps of: receiving (602), via an input device, a natural language question statement (402) from a user, the question statement indicating questions that have occurred during operation of the packaging line (1); generating (604), by a trained language model (42), one or more instructions (404) based at least in part on the question statement (402), the one or more instructions (404) indicating an operation to be executed by a user on the packaging line (1) to alleviate the question; and outputting (606), by an output device, the one or more instructions (404) in a natural language to a user. A method (500) of training a language model (42), a language model (42), a data processing device (4), a packaging line (1) and a computer program are also disclosed.

Description

Technical Field

[0001] This invention relates to the field of packaging production technology, and in particular to a computer-aided troubleshooting method for a packaging production line configured to produce packaging, a method for training a language model for the method, and a corresponding language model, data processing device and computer program. Background Technology

[0002] As is well known, many liquid, semi-liquid, or pourable foods, such as fruit juice, UHT (ultra-high temperature sterilized) milk, wine, and ketchup, are distributed and sold through packaging, especially composite packaging made of multi-layer composite packaging materials.

[0003] A typical example is called Tetra Brik Aseptic TM The packaging material is a parallelepiped-shaped package for pourable food, made by sealing and folding laminated strip-shaped packaging material. This packaging material has a multi-layered structure comprising cardboard and / or paper base layers, both sides of which are covered with heat-sealable plastic layers, such as polyethylene. For aseptic packaging of products intended for long-term storage, the packaging material also includes layers of oxygen-barrier material, such as aluminum foil, stacked on top of the heat-sealable plastic layers, and then covered with another heat-sealable plastic layer, which forms the inner surface of the package that ultimately contacts the food.

[0004] These types of composite packaging are typically produced on packaging production lines, also known as packaging lines or filling lines. These lines can operate fully or nearly fully automatically and can begin by forming a package from a roll of multi-layer composite packaging material (wound from a spool) and then filling the composite package with pourable food.

[0005] A typical packaging line includes at least one filling machine that forms composite packages using multi-layer composite packaging materials and fills the composite packages with pourable food. Additionally, the packaging line may include downstream processing equipment that receives the composite packages from the filling machine and performs further processing on them; the downstream processing equipment may include, for example, one or more buffer units to temporarily cushion the composite packages; an application unit for applying, for example, straws, to the composite packages; and a grouping unit, such as a palletizing unit, for grouping multiple composite packages into storage units (e.g., trays).

[0006] Specifically, multi-layer composite packaging materials are typically supplied in roll form, particularly wound on a packaging material reel, and then fed into a filling machine. In the filling machine, the roll is formed into a tubular shape by longitudinal sealing. Liquid food is fed into the tube through the tube; then the lower end of the tube is fed into a folding device, where it is transversely sealed. The tube is folded along the fold line (also known as the weakening line) and then cut, thus forming a composite package containing the liquid food.

[0007] In such packaging lines, especially in the production of tiltable food composite packaging, defect monitoring and detection are crucial due to the extremely high speeds of the lines used to increase output. At the same time, the technical complexity of these packaging lines makes it difficult for operators to identify the causes of machine malfunctions and appropriate solutions; solutions must be found very quickly to avoid production interruptions or negative impacts.

[0008] Therefore, the purpose of this disclosure is to provide technology to help packaging line operators quickly find key answers when machine malfunctions occur during packaging line operation. Summary of the Invention

[0009] This objective is addressed by the subject matter defined in the independent claims. Beneficial modifications to embodiments of this disclosure are defined in the dependent claims, as well as in the specification and drawings.

[0010] One aspect of this disclosure relates to a method for computer-aided troubleshooting of packaging lines configured to produce packaging. Troubleshooting packaging production lines is a significant and challenging technical task for a variety of reasons. First, packaging lines are a critical component of manufacturing operations. When a packaging line encounters problems or malfunctions, it can disrupt the entire production process, leading to delays, reduced productivity, and increased downtime. Troubleshooting is essential for timely identification and resolution of problems to minimize the impact on production efficiency. However, packaging lines typically contain numerous interconnected components, such as conveyors, sensors, filling machines, sealing machines, labeling machines, and / or control systems. The complexity and integration of these components make troubleshooting extremely challenging. Identifying the root cause of a problem requires a deep understanding of the entire system and the individual components and their interactions. Effective packaging line troubleshooting may require knowledge of mechanical systems, circuits, control systems, automation, sensor technology, programming, and more. Furthermore, troubleshooting often needs to be performed in real time, or at least as quickly as possible, to minimize downtime and production losses. This necessitates rapid decision-making and timely implementation of corrective actions. Packaging lines often operate in challenging environments, such as high-speed production lines. This could affect the performance and reliability of the equipment, making troubleshooting more complex.

[0011] According to one aspect of this disclosure, the method may include the step of receiving a natural language problem statement from a user via an input device. The problem statement may indicate a problem occurring during the operation of a packaging line. The method may include the step of generating one or more instructions, at least in part, based on the problem statement, which instruct actions to be performed by the user on the packaging line to alleviate the problem. The step of generating one or more instructions may be performed by a trained language model. The method may include the step of outputting one or more instructions to the user in natural language via an output device.

[0012] This provides a computer-implemented assistant or agent that can interact in a conversational manner to help users obtain relevant information about the packaging line, particularly for troubleshooting purposes. The natural language conversational format makes it possible to answer questions, infer text from user manuals, and / or provide various useful information (such as charts and / or images).

[0013] The methods disclosed herein can assist users in performing technical tasks related to troubleshooting packaging lines through a continuous and / or guided human-computer interaction process. Specifically, one or more instructions provided by the methods disclosed herein may relate to internal states prevalent in the packaging line and / or its components (see examples below), thereby enabling users to operate the packaging line correctly.

[0014] The method disclosed herein is preferably implemented during the actual operation of existing packaging lines, especially when users discover problems in the operation of the packaging lines. Therefore, the method disclosed herein can be understood as being embedded in the overall operation of the packaging line.

[0015] When the output of the method (i.e., one or more instructions generated by the language model) is output to the user (the same user who provides the question statement received through the input device) via the output device, this creates a closed-loop process for human-computer interaction.

[0016] As described above, one or more instructions direct the user to perform actions on the packaging line to alleviate the problem. Therefore, in the intended use of the disclosed method, the user will normally follow one or more instructions generated by the language model, thus essentially making the user a link in the technology chain, such that one or more instructions, and therefore the output of the language model, are directly related to physical reality, or at least have implicit technical effects. The language model itself represents a technical tool within the disclosed method because it is applied in the technical field (i.e., packaging production) and used to solve a specific technical problem, namely, assisting in troubleshooting the packaging production line.

[0017] A packaging production line should be understood as a technical facility configured to produce packaging (particularly food packaging), specifically liquid, semi-liquid, or pourable foods, such as those mentioned above. A packaging line may include at least one filling machine configured to form packaging from multi-layer composite packaging materials and fill the packaging with product. The packaging line may also include downstream processing equipment configured to receive composite packaging from the filling machine and perform additional processing on the packaging. Downstream processing equipment may include, but is not limited to: one or more cushioning units for temporary cushioning of the packaging, application units for applying, for example, straws to the packaging, and / or grouping units (e.g., palletizing units) for grouping multiple packages into storage units (e.g., trays). Packaging lines, the subject of the methods disclosed herein, may be arranged in a manufacturing plant or system.

[0018] Users should be understood as individuals who work on or otherwise interact with or are associated with the packaging line. For example, a user could be a packaging line operator, technician, or maintenance personnel.

[0019] Problem statements in natural language should be understood as part of a user-provided language used to describe problems that arise during the operation of the packaging line. The methods disclosed herein may not impose material restrictions on the format and / or content of problem statements, or may not impose any restrictions at all. Instead, users are free to formulate problem statements.

[0020] Non-limiting examples of problem statements may include:

[0021] The top of package number 20 has a dent.

[0022] Inconsistent changes in the liquid level within the packaging can lead to insufficient (or excessive) filling.

[0023] Leaks or spills frequently occur during the filling (or sealing) process at packaging lines.

[0024] The packaging was not sealed properly.

[0025] The labeling process can cause labels on packaging to become misaligned (or unreadable).

[0026] The packaging line is blocked or clogged.

[0027] The packaging line is unable to maintain the required production speed.

[0028] The barcode or QR code on the packaging cannot be scanned (or contains error information).

[0029] The packaging line produces packages with inconsistent (or inaccurate) product codes (or batch numbers).

[0030] The packaging materials (such as bottle caps or seals) are defective or cannot securely seal the packaging.

[0031] The packaging line was down for too long due to equipment malfunction or failure.

[0032] The packaging line produces packages with damaged or scratched surfaces.

[0033] The packaging gets stuck or misaligned on the conveyor system.

[0034] Excessive foaming during the filling process of liquid products can lead to spillage or packaging integrity issues.

[0035] The packaging line produces packages with inconsistent or incorrect label information (such as ingredients or nutritional information).

[0036] The packaging is subjected to excessive vibration or shaking during the packaging process, causing the product to agitate or bubble.

[0037] The packaging line failed to maintain the temperature or environmental conditions required for the product.

[0038] The packaging line has difficulty handling packages of different sizes or specifications.

[0039] The packaging line produces packaging with illegible or stained labels or packaging materials.

[0040] The packaging fell or was knocked over during the packaging process.

[0041] One or more instructions that direct the user to perform actions on the packaging line to alleviate a problem should generally be understood as any information that may help or assist the user in resolving the problem stated in the statement.

[0042] One or more instructions may contain specific actions recommended for the user to perform. Non-limiting examples of instructions relating to the exemplary problem statement above may include:

[0043] Try setting the fill pressure higher (this instruction can also explain the possible causes of the problem to enhance the user's understanding, such as "the dent may be caused by insufficient fill pressure, try setting it higher").

[0044] Adjust filling settings or parameters to ensure accurate and consistent filling levels. Calibrate or replace any faulty sensors or devices involved in the filling process.

[0045] Inspect and repair any faulty valves, seals, or connections. Ensure that all components involved in liquid delivery and sealing are properly aligned and tightened.

[0046] Verify that the sealing equipment is functioning correctly and that the appropriate temperature and pressure are set. Replace worn or damaged sealing components, such as heating elements or sealing strips.

[0047] Inspect and adjust the label applicator to ensure proper alignment and accuracy. Verify the quality and position of the label roll. Ensure the correct label format and print settings are used.

[0048] Inspect and clean the conveyor system, removing any debris or obstructions. Ensure the conveyor belt is properly aligned and tensioned. If necessary, check and adjust the guide rails or partitions.

[0049] Identify and resolve bottlenecks or issues in the packaging production line that are slowing down production. Optimize equipment setup, streamline processes, or consider upgrading equipment where feasible.

[0050] Ensure the barcode or QR code printing settings are correct and the print quality is high. Verify the readability of the code using appropriate scanning equipment. Replace the faulty or damaged code printer if necessary.

[0051] Inspect the product's coding or labeling system. Inspect and calibrate the inkjet printing equipment. Verify the accuracy of data input or integration with the inkjet printing system.

[0052] Inspect and replace any defective or damaged packaging materials, such as bottle caps or seals. Adjust equipment settings to ensure the packaging is properly sealed and secure.

[0053] Follow the preventative maintenance plan to ensure equipment is regularly inspected, lubricated, and maintained. Report any malfunctions to the maintenance or engineering team promptly for repair or adjustment.

[0054] Inspect and replace any worn or damaged components involved in the packaging process, such as guide rails or supports. Verify that protective elements (such as brushes or dials) are in the correct position and function.

[0055] Inspect and adjust the conveyor system to ensure proper alignment and tension. Inspect and repair any faulty or misaligned conveyor belts, rollers, or guide rails.

[0056] Optimize filling parameters to minimize foaming. Consider adjusting the liquid flow rate. Ensure the filling equipment (including nozzles or dispensing heads) is functioning properly.

[0057] Review and verify the label data source or system. Audit the label printing and inspection process.

[0058] Inspect and balance moving components, such as conveyors or rotating components. Evaluate the installation or assembly of packaging equipment. Consider implementing shock absorption measures or vibration isolation.

[0059] Monitor and adjust temperature and environmental conditions within the packaging area. Ensure that the insulation, ventilation, or heating / cooling systems required for liquid products are adequate.

[0060] Streamline conversion procedures and minimize downtime. Provide operators with clear instructions or visual aids to efficiently switch between packages of different sizes or specifications. Consider implementing standardized conversion kits or quick-change components.

[0061] Verify the print quality and settings of the printing equipment. Inspect and replace ink cartridges or printheads as needed. Ensure the printing system receives proper maintenance and cleaning.

[0062] Review and optimize the design or configuration of the packaging line to minimize packaging errors. Inspect and adjust conveyor speeds or guide rails to ensure proper packaging handling and stability.

[0063] In addition to specific recommended actions, one or more instructions may contain other information and may be used in conjunction with one or more specific recommended actions. For example, one or more actions may include documents, charts, graphs, photographs, illustrations, lists, historical process data, production logs, quality control reports, maintenance records, tables, flowcharts, decision trees (or links to these elements), or any other information that may help or assist the user in resolving the problem stated in the problem statement. Thus, the language model that generates one or more instructions can be understood as a retrieval-enhanced language model or a retrieval-enhanced LLM.

[0064] Input devices for receiving question statements can include any technological device configured to capture input in natural language provided by the user as part of the claimed method. Different input modes or combinations thereof may be supported. A keyboard (e.g., a traditional computer keyboard embodied as a physical keyboard or a virtual keyboard displayed on a touchscreen) is one example of an input device that allows a user to input a question statement in text form. Another example is a microphone that allows a user to input a question statement orally. Spoken language can be converted to text using a speech-to-text algorithm, and then processed by a language model. Another example of an input device is a pen-based input device, such as a digital stylus or writing tablet, which can be used to write or draw natural language characters on a touch-sensitive surface.

[0065] Output devices used to deliver one or more instructions to a user can include any technology suitable for outputting natural language to a user. Different output modes or combinations thereof can be supported. For example, a display device (such as a computer screen or mobile device screen) can present natural language output, displaying one or more instructions on the screen for the user to read, and allowing the user to visually inspect images, graphics, charts, etc., contained within one or more instructions. As another example, a speaker can convey one or more instructions in an audible form. Text-to-speech (TTS) technology can be used to convert written text into synthesized speech, enabling the user to hear one or more instructions.

[0066] Input and / or output devices can be directly mounted on the packaging line. Alternatively, input and / or output devices can be part of a device associated with and / or connectable to the packaging line, such as data processing equipment (e.g., control computers, workstations, terminals, etc.), mobile devices (e.g., smartphones, tablets, etc.), or wearable devices (e.g., virtual reality and / or augmented reality devices, smart glasses, smartwatches, etc.). Input and output devices can be integrated into the same device.

[0067] A language model should be understood as a computer model configured to understand natural language provided as input and generate natural language as output. Language models can be configured to recognize, summarize, translate, predict, and / or generate text and / or other forms of content. Examples of language models include, but are not limited to, OpenAI's GPT-3 and GPT-4, Meta's LLaMA, and Google's PaLM2. Language models can be based on or incorporate principles from computer science, particularly those related to Natural Language Processing (NLP) and machine learning.

[0068] In the context of NLP, language models are typically designed to capture the complex patterns, structures, and / or semantics of human language. They can utilize computational algorithms and / or statistical models to learn the relationships between words, phrases, and / or sentences in a given language. By analyzing large amounts of text data, language models can gain the ability to make predictions, generate coherent text, and / or perform various language-related tasks.

[0069] Machine learning refers to the algorithms and statistical models used by computer systems to perform specific tasks, without the need for explicit instructions, but relying on models and inference. For example, in machine learning, instead of using rule-based data transformations, data transformations inferred by analyzing historical and / or training data can be used.

[0070] Language models can be deep learning models. A language model can contain or consist of an artificial neural network with multiple parameters. At its core, a language model is typically implemented in the form of a recurrent neural network (RNN) and / or a Transformer architecture. These architectures enable language models to process sequential information and capture dependencies between words and phrases. The parameters of a language model (also known as weights) can be learned through a process called training, which involves feeding the model a large amount of textual data and adjusting the parameters based on observed data patterns and statistical properties. The training of a language model will be described in more detail below.

[0071] Language models can contain a large number of parameters, often ranging from millions to tens of millions to billions, hence such models are often referred to as large language models (LLMs). Although language models are typically trained on relatively simple tasks (such as predicting the next word in a sentence), it has been found that well-trained and / or parameter-counted language models can adequately capture most of the syntax and semantics of human language.

[0072] According to another aspect of this disclosure, the language model has been trained using at least one operation manual related to the packaging line. This operation manual may be a comprehensive document providing detailed instructions, guidelines, and / or information on how to properly operate, maintain, troubleshoot, and / or perform routine tasks on the packaging line. The operation manual can serve as a reference guide for users. It may contain many pages, typically at least several hundred pages, and sometimes even exceeding 2,000 pages. The operation manual may be provided in an electronic document format (e.g., PDF).

[0073] The user manual may contain the following sections or any combination thereof:

[0074] Introduction: Overview, uses, and / or key characteristics of the packaging line. This section may also provide an introduction to the manual itself and an overview of its organization.

[0075] Safety Guidelines: Detailed safety instructions and / or precautions that must be followed to ensure operator safety and prevent accidents. This section may cover topics such as the proper use of personal protective equipment (PPE), emergency procedures, and / or specific safety considerations related to packaging line operations.

[0076] Machine Description: A comprehensive description of the packaging line, including its various components, controls, interfaces, and / or functions. This section may include diagrams, labeled illustrations, and / or photographs to aid in understanding the structure and / or layout of the packaging line.

[0077] Operating Procedures: Step-by-step instructions on how to set up, configure, and / or operate the packaging line. This section may cover tasks such as start-up and / or shutdown procedures, material loading, setup adjustments, running production cycles, and / or monitoring machine performance.

[0078] Maintenance and Troubleshooting: A guide to routine maintenance tasks, including cleaning, lubrication, and / or periodic inspections. This section may also provide troubleshooting procedures to identify and / or resolve common problems and / or malfunctions that may occur during operation. This section may contain diagnostic charts, error code descriptions, and / or recommended corrective actions. This section aims to ensure optimal performance, reliability, and / or service life of the packaging line. This section may typically contain the following subtopics or combinations thereof:

[0079] a) Routine Maintenance: Description of routine maintenance tasks, including regular cleaning, lubrication, and / or inspection procedures. This section may specify recommended cleaning agents, methods, and / or frequencies for cleaning different machine components. Lubrication guidelines may indicate the specific lubricant to use, its application points, and / or recommended lubrication schedules. Routine inspections may involve checking for loose or damaged components, worn components, and / or signs of wear.

[0080] b) Preventive Maintenance: This section may focus on proactive measures to prevent potential problems and maximize the lifespan of the packaging line. It may include a plan for preventive maintenance tasks, such as replacing filters, belts, and / or other wear-prone components. Additionally, it may provide guidelines for calibrating sensors, adjusting settings, and / or verifying measurement accuracy to ensure consistent performance.

[0081] c) Troubleshooting Procedures: Detailed troubleshooting procedures are provided to help users identify and resolve common problems encountered during machine operation. This may include a comprehensive list of potential problems, possibly accompanied by step-by-step instructions for diagnosing the causes and / or taking appropriate corrective actions. Troubleshooting diagrams, flowcharts, and / or decision trees may be included to help systematically resolve problems. Each problem may have a series of corresponding potential causes, symptoms, and / or recommended solutions.

[0082] d) Error Codes and Alarms: This section may cover the error codes or alarm systems for the packaging line and their meanings. It may provide a complete list of error codes, alarms, and / or warnings that may be displayed on the control panel or interface. For each code or alarm, an explanation may be provided to help users identify the problem and take necessary corrective action. It may also include instructions on how to clear or reset the error code after resolving the underlying problem.

[0083] e) Maintenance Tools and Spare Parts: This section may provide information on the specific tools, equipment, and / or consumables required for maintenance tasks. It may list recommended tools, their uses, and / or instructions on how to use them effectively. Additionally, there may be a dedicated spare parts section specifying part numbers, instructions, and / or recommended sources for obtaining replacement components.

[0084] f) Contact Information and Technical Support: The user manual may contain contact information for technical support, customer service, and / or the manufacturer's service department. This information helps users seek further assistance when encountering complex problems or questions beyond the scope of the manual. It may include telephone numbers, email addresses, and / or website links to online resources or knowledge bases.

[0085] Technical Specifications: Detailed specifications of the packaging line and / or its components, such as power requirements, dimensions, weight, capacity, speed and / or any other relevant technical information.

[0086] Appendix: Other reference materials, such as electrical diagrams, spare parts lists and / or recommended suppliers of replacement components.

[0087] Training with at least one instruction manual and / or any of the aforementioned information related to the packing line helps to create a language model that possesses a comprehensive knowledge of the packing line's target parameters, conditions, and / or states, enabling the language model to answer a wide range of question statements.

[0088] According to another aspect of this disclosure, the language model has been trained using process data collected from the packaging line. The process data may contain information about the packaging line, information about one or more components of the packaging line, and / or information about the packaging produced. The process data may be generated during the operation of the packaging line. The process data may be collected using sensors; that is, it is measurement data reflecting the technical characteristics of the packaging line, one or more components of the packaging line, and / or the packaging produced. Specific process data may vary depending on the complexity and capacity of the packaging line and the type of product being packaged. Examples of process data collected in the packaging line may include any of the following:

[0089] Production metrics: Data related to the production output and / or efficiency of the packaging line. This may include the number of units packaged per unit time (e.g., packages per minute), Overall Equipment Effectiveness (OEE), productivity, downtime, and / or overall production line efficiency.

[0090] Machine performance data: Data relating to the performance of individual machines or components within the packaging line. This may include data on machine speed, cycle time, operating temperature, pressure, power consumption, and / or motor or drive performance.

[0091] Quality control parameters: Data related to quality control checks and / or measurements performed during the packaging process. This may include data on weight checks, visual inspections, barcode scanning, fill levels, seal integrity, label accuracy, and / or other quality attributes specific to the packaged product.

[0092] Material usage and inventory: Data related to the usage of packaging materials, such as consumption, material waste or scrap rate, roll or reel replacement, and / or inventory levels of packaging materials or consumables.

[0093] Maintenance and repair data: Data related to maintenance activities, repairs, and / or adjustments performed on the packaging line. This may include maintenance plans, task completion records, parts replacement history, and / or equipment downtime due to maintenance or repair.

[0094] Fault and Alarm Log: Records data on any faults, errors, or alarms encountered during the packaging process. This may include error codes, alarm notifications, fault descriptions, timestamps of occurrence, and / or any actions taken to resolve the issue.

[0095] Environmental conditions: Data relating to environmental conditions in and / or within the packaging line area, such as temperature, humidity, air quality and / or other factors that may affect the packaging process or product quality.

[0096] Operator logs and feedback: Data recorded by users, such as shift logs, operator feedback, observations and / or records of machine performance, process improvements and / or potential problems encountered during operation.

[0097] Training with process data collected from the packaging line helps create a language model that possesses a comprehensive knowledge of the actual parameters, conditions, and / or states of the existing packaging line. This knowledge may differ from what is specified in the operator's manual, depending on the specific setup of the existing packaging line, enabling the language model to answer a wide range of question statements.

[0098] Specifically, if both the collected process data and at least one of the operation manuals mentioned above are used for training, the model can be further fine-tuned, and the likelihood of providing more accurate answers to the problem statements will be further increased.

[0099] According to another aspect of this disclosure, the language model has been trained using stored information indicating previously encountered problems and one or more instructions, corresponding to previously generated instructions, to be performed to alleviate those problems. Therefore, a feedback loop can be provided to (re)train the language model using the problems that have occurred and / or the corresponding recommended instructions. In this way, the language model can improve over time during the wrapper line operation, thereby further enhancing its accuracy.

[0100] According to another aspect of this disclosure, the method may include a step of validating one or more instructions generated by the language model before incorporating them into the training of the language model. This validation can be performed by a user. This ensures that only accurate answers to the question statement are incorporated into the training and ultimately recommended by the language model.

[0101] According to another aspect of this disclosure, the step of generating one or more instructions is also based, at least in part, on process data collected from the packaging line. Therefore, the language model takes both the problem statement and the process data as input. The process data used for this purpose may include any process data further described above regarding language model training. This allows for the execution of steps to determine whether a given measurement's process data value equals a target value, reaches a target threshold, or falls within a target range. The language model may perform this step during the generation of one or more instructions. The target value, target threshold, and / or target range may be predefined or derived by the language model, for example, based on information from at least one operating manual associated with the packaging line.

[0102] According to another aspect of this disclosure, the step of generating one or more instructions includes generating multiple instructions, wherein the step of outputting includes outputting the multiple instructions as a sequence of operations to be performed by the user. Therefore, the language model can provide the user with particularly detailed and comprehensive recommendations, for example, in the form of a step-by-step checklist.

[0103] According to another aspect of this disclosure, the step of generating one or more instructions further includes executing one or more dialogue cycles in response to receiving a question statement. One or more dialogue cycles may include generating a specific request in natural language from a trained language model to prompt the user for additional information relevant to the question. One or more dialogue cycles may include outputting the specific request in natural language to the user via an output device. One or more dialogue cycles may include receiving additional information relevant to the question in natural language from the user via an input device. One or more dialogue cycles may include generating one or more candidate instructions from a trained language model, instructing the user on an action to be taken to alleviate the problem. One or more dialogue cycles may continue to execute until one or more candidate instructions reach a predetermined confidence score. The confidence score can be manually set according to the specific use case. The output step may include outputting one or more candidate instructions to the user in natural language. Therefore, if the original question statement is too vague to generate a suitable answer, the language model can request additional or more specific information from the user. In this way, the conversational capabilities of the language model can be fully utilized, providing advantages such as improved context awareness, enhanced coherence, and a more human-like dialogue flow. This allows language models to better understand user intent, recall prior information, and generate responses built upon the previous context, thus providing a specially guided human-computer interaction process.

[0104] According to another aspect of this disclosure, the steps of performing one or more dialogue cycles are based on thought chain (CoT) cues. CoT cues should be understood as a cuing method that encourages language models to interpret their reasoning. CoT cues generally refer to a technique that facilitates the generation of coherent and / or context-relevant language output based on a series of cues or clues. It aims to enhance the language model's ability to maintain the coherence and / or logic of information flow throughout the dialogue or text generation process. In traditional language models, each sentence or response is typically generated independently, without regard to context or previous statements, which can lead to disjointed or inconsistent output. With CoT cues, language models can be configured to incorporate contextual information from previous dialogue cycles to generate more coherent and relevant responses. CoT cues may involve capturing and utilizing contextual information from previous dialogue cycles to influence the generation of subsequent language output. By maintaining the "thought chain," the language model can better understand the intended meaning and continue the dialogue in a logical and coherent manner. CoT cues can be implemented using various mechanisms. One approach is to employ recurrent neural networks (RNNs) or Transformer-based models, which are capable of preserving and utilizing information from previous steps or tokens. These models can be trained to learn the dependencies between cues and generate responses that correspond to the ongoing dialogue.

[0105] According to another aspect of this disclosure, the method further includes the step of generating at least one control signal. This step may be performed at least in part based on one or more instructions. This step may be performed by a trained language model. Alternatively, this step may be performed by a software component that receives the output of the language model (i.e., one or more instructions). The control signal may be configured to control and / or modify the operation of the wrapper line to alleviate the problem.

[0106] Non-limiting examples of control signals include:

[0107] Start / Stop Signals: These signals can start or stop the packaging line. By sending a stop signal, users can quickly resolve any issues or perform maintenance tasks, thus preventing further problems from occurring.

[0108] Speed ​​control signals: Adjusting the speed helps regulate conveyors, robotic arms, or other moving components on the packaging line. Decreasing or increasing packaging line speed helps prevent product blockages, reduce errors, or meet different packaging requirements.

[0109] Synchronization signals: As mentioned above, packaging lines typically consist of multiple components, machines, and / or stations working in concert. Synchronization signals ensure that the different components of the packaging line operate in a coordinated manner. For example, a filling machine can send a signal to a capping machine to indicate that the container is ready for capping.

[0110] Quality control signals: These signals can be used to monitor and / or maintain the quality of packaged products. For example, vision systems or sensors can check the size, color, or barcode of a product and send signals to reject any products that do not meet specified standards.

[0111] Maintenance signals: Packaging lines typically require regular maintenance. Maintenance signals can be sent to indicate when specific machines and / or components need servicing, such as lubrication, calibration, or parts replacement.

[0112] The generated control signal may directly affect the operation of the packaging line. In this case, the disclosed method may include steps of automatically controlling and / or modifying the operation of the packaging line based on the at least one control signal. Alternatively, the at least one control signal may be executed only after confirmation by the user (e.g., the user who provided the original problem statement).

[0113] According to another aspect of this disclosure, the language model has been trained using self-supervised training. Self-supervised training should be understood as a machine learning technique that allows the language model to learn from unlabeled data by developing its own learning tasks. Unlike supervised learning, which requires labeled examples, self-supervised training can automatically create surrogate or “pseudo” labels by leveraging the inherent structure or patterns in the data.

[0114] In the context of training language models, self-supervised learning typically involves predicting missing or masked portions of input text. The model might be fed a corrupted version of a sentence where parts of the text have been replaced with special lexical units, such as the [MASK] lexical unit. The goal of the model is to predict the original content of the masked portion based on the context provided by surrounding words.

[0115] By training a language model using self-supervised learning, the language model can learn rich contextual representations of words and sentences, thereby generating coherent and context-sensitive responses.

[0116] According to another aspect of this disclosure, self-supervised training may include some, all, or any suitable combination of the following steps to train a language model:

[0117] Dataset Preparation: Collect training data, typically a large amount. Training data can be collected from any of the sources described in this paper, such as at least one instruction manual related to the wrapper line. This data serves as the training corpus for the language model.

[0118] Preprocessing: This can be done by breaking down the training data primitives into smaller units (such as words or subwords) and converting them into numerical representations suitable for model training.

[0119] Masking: During training, a certain percentage of words in the input text may be randomly masked or replaced with [MASK] words. The goal of the model is to predict the original content of the masked words.

[0120] Model architecture: Initialize the language model architecture, such as a Transformer-based model. Transformers are well-suited for capturing contextual information from surrounding words, making them highly effective in self-supervised learning tasks.

[0121] Training: The language model is trained using a masked input sentence. The model predicts the original content of the masked lemmas by considering the context of surrounding words. The model's parameters are updated using an optimization algorithm (e.g., stochastic gradient descent (SGD)) to minimize the prediction error.

[0122] Iterative training: The training process can be repeated multiple times, during which the model is exposed to different masked instances of the input sentence. This can help the model learn increasingly complex patterns and dependencies in the language.

[0123] Fine-tuning: After initial self-supervised training, the model can be further fine-tuned for specific downstream tasks (such as text classification, language translation, or question answering). This fine-tuning process can help adapt the representations learned by the model to the current task, thereby improving its performance on that specific task.

[0124] The above random masking may involve randomly selecting a portion of the input tokens and replacing them with a special [MASK] token. This process may involve some or all of the following steps:

[0125] Masking probability: You can define a masking probability or percentage, indicating the approximate percentage of words in the input text that will be randomly masked. For example, if the masking probability is set to 15%, approximately 15% of the words will be masked.

[0126] Random selection: Then, randomly selected words are chosen for masking according to the defined masking probability. The selection is usually uniform or follows a random distribution. For example, if the sentence contains 10 words and the masking probability is 15%, approximately 2 words will be selected for masking.

[0127] Masking operation: The selected word is replaced with a special "masking" word, denoted as [MASK]. This replacement hides the original content of the word, thus creating gaps in the input sequence.

[0128] Training objective: The goal of the language model during training is to predict the original content of the masked lexical units based on the context provided by the surrounding words.

[0129] By randomly masking parts of the input vocabulary, the language model is exposed to diverse and partially corrupted examples, forcing it to learn robust representations that capture latent patterns and dependencies in the text. This helps the language model gain a deeper understanding of language and its context, thereby improving its language generation and comprehension capabilities.

[0130] According to another aspect of this disclosure, the language model has been validated using one or more first instructions generated by the language model and one or more second instructions generated by a human user, wherein the one or more first and second instructions are generated in response to the same problem statement or a set of problem statements. Therefore, the one or more artificially generated second instructions serve as a basis for validating the language model. In one implementation, the generated one or more first and second instructions can be presented to a second user who can compare the instructions and determine which better addresses the problem statement. Validation can be performed during or after training.

[0131] According to another aspect of this disclosure, the language model includes neural networks, particularly deep neural networks and / or Transformer architectures. When a language model is implemented based on a Transformer architecture, the language model may contain one or more of the following components:

[0132] Input embedding components: Input embedding components can be used to segment input text into smaller units, such as words or subwords. Each lexical unit can be converted into a numerical representation called an embedding. These embeddings can capture the semantic and contextual information of each lexical unit.

[0133] Positional Encoding Component: The positional encoding component can be configured to provide positional information to the language model. For this purpose, positional encodings can be added to the lexical embeddings. These encodings can indicate the position of each lexical in the input sequence and can help the language model understand the order and / or relative position of the lexical units.

[0134] One or more encoder layers: A Transformer-based language model can contain stacked identical encoder layers or consist of a set of identical encoder layers. Each encoder layer can contain two sub-layers or consist of two sub-layers.

[0135] a) Self-attention mechanism: The self-attention mechanism allows language models to weigh the importance of different words in the input sequence when generating representations. It can capture the dependencies between different positions in the input sequence. The self-attention mechanism can calculate attention scores between all pairs of input positions and generate weighted representations by focusing on relevant words.

[0136] b) Feedforward Neural Network: After the self-attention sublayer, a feedforward neural network can be applied independently to each position. This network can contain two linear transformations and a nonlinear activation function in between, or be composed of both, enabling the language model to capture complex relationships between words.

[0137] One or more decoding layers: In addition to the encoder layer, Transformer-based language models can also contain stacked decoding layers. The decoding layers can generate output tokens based on representations learned from the encoder layer. A decoding layer can also contain two sub-layers or consist of two sub-layers:

[0138] a) Masked Self-Attention: In the decoding layer, the self-attention mechanism can be masked to prevent attention from being paid to subsequent positions. This masking ensures that the language model can only access information from previously generated lexical units during the decoding process.

[0139] b) Encoder-decoder attention mechanism: The decoder layer may also incorporate an attention mechanism into the encoder layer representation. This allows the language model to utilize the contextual information from the encoder when generating output lexical units.

[0140] Output layer: The output layer may contain linear transformations. These linear transformations may be followed by a softmax activation function. It maps the final representation to a probability distribution of the vocabulary, enabling the language model to generate the most probable next word in the sequence.

[0141] The Transformer architecture enables language models to effectively capture long-range dependencies in input sequences through self-attention mechanisms. Transformer-based language models can process input sequences in parallel, making them highly scalable and well-suited for training on large amounts of text data. This parallelism reduces training and inference time, further enhancing the scalability of Transformer-based models and their suitability for handling large volumes of text data. More specifically, Transformer-based language models can provide one or more of the following technical benefits:

[0142] Reduce processing resources:

[0143] a) Parallel processing: The Transformer can process input sequences in parallel, thus making efficient use of hardware resources such as CPUs and GPUs. This parallelism can accelerate the training and / or inference process and reduce overall computational requirements.

[0144] b) Optimized Architecture: Transformers are designed to improve computational efficiency. They typically use self-attention mechanisms, which, compared to sequential models such as recurrent neural networks (RNNs), are able to capture long-range dependencies with less computation. Therefore, Transformer-based models can achieve comparable or better performance using fewer processing resources.

[0145] Storage space:

[0146] a) Embedding size: Compared to traditional methods such as one-hot encoding, Transformer-based models typically use low-dimensional embeddings. This reduction in dimensionality can significantly reduce the storage space required to store model parameters.

[0147] b) Parameter sharing: The Transformer architecture can employ parameter sharing, meaning that the same parameters are used at different positions in the input sequence. This sharing reduces the number of unique parameters that need to be stored, thereby reducing storage requirements.

[0148] Network bandwidth:

[0149] a) Model compression: Transformer-based models can be compressed using various techniques, such as quantization, pruning, or knowledge distillation. These methods can reduce model size without significantly sacrificing performance, thereby reducing network bandwidth consumption during model deployment and / or inference.

[0150] b) Device-side execution: Transformers can be deployed on edge devices, such as smartphones or IoT devices, and can execute on the device without relying on cloud-based services. This reduces the need for frequent network data transmission, thereby minimizing network bandwidth consumption.

[0151] Pre-trained models and transfer learning:

[0152] a) Model Distribution: Pre-trained Transformer models can be distributed, allowing users to download and use them without having to train them from scratch. This distribution method reduces the need for individual users to train large models, thereby saving processing resources and reducing network bandwidth consumption.

[0153] b) Fine-tuning: Through transfer learning, pre-trained Transformer models can be fine-tuned for specific downstream tasks. This approach allows pre-trained models to be reused and adapted to new tasks, thereby reducing the required training time and resource consumption.

[0154] By leveraging the advantages of parallel processing, optimized architecture, model compression techniques, and / or transfer learning, Transformer-based language models can reduce processing resource, storage space requirements, and / or network bandwidth consumption. These effects are particularly beneficial for deployment in resource-constrained environments, such as the industrial production environment described in this disclosure, where efficient use of computing resources and network bandwidth is crucial.

[0155] However, it should be understood that the concepts of this disclosure are not strictly limited to Transformer-based language models. Instead, language models can be implemented using alternative techniques, such as:

[0156] Recurrent Neural Networks (RNNs), such as Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), can process sequences sequentially and maintain internal hidden states to capture context. RNNs can be particularly effective at handling sequential data and short-term dependencies. However, RNNs may struggle to capture long-range dependencies due to vanishing or exploding gradients.

[0157] Convolutional Neural Networks (CNNs): CNNs are primarily used for computer vision tasks, but can also be applied to language modeling. CNNs utilize one or more convolutional layers to capture local patterns and dependencies in an input sequence. CNNs are computationally efficient and capable of capturing hierarchical structures in data. However, compared to RNNs or Transformers, CNNs may have limited ability to model long-term dependencies and sequence relationships.

[0158] Recurrent Convolutional Neural Networks (RCNN): RCNN combines the advantages of RNNs and CNNs. It can combine convolutional layers to capture local patterns and / or recurrent connections to capture long-term dependencies. RCNNs perform well in tasks such as sentiment analysis, text classification, and machine translation. However, compared to Transformers, they still face challenges in modeling complex language structures and dependencies.

[0159] Gated Convolutional Neural Networks (GCNNs): GCNNs may aim to address the limitations of CNNs by introducing gating mechanisms similar to those in RNNs. These gating mechanisms can help control the flow of information and / or enable the model to capture long-range dependencies. GCNNs have been used in tasks such as machine translation and language generation and have shown good results. However, in some cases, they may be less efficient at capturing dependencies than Transformers.

[0160] Hybrid architectures combine multiple neural network components, such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), or transformers, to leverage their respective strengths. For example, a language model can combine CNNs for local pattern extraction and / or transformers to capture global context and / or long-range dependencies. Hybrid architectures can strike a balance between computational efficiency and capturing complex language structures.

[0161] According to another aspect of this disclosure, the language model is deployed on a container management platform, such as Kubernetes. A container management platform (also known as a container orchestration platform) should be understood as a software framework configured to automate the deployment, scaling, and / or management of containerized applications. It can provide a lightweight and / or isolated environment for packaging applications and / or their dependencies, thereby enabling easy portability and scalability. Deploying the language model on a container management platform offers several technical advantages, including:

[0162] Scalability: Container management platforms excel at horizontally scaling applications. When implementing language models within a container management platform, the number of model instances can be easily scaled up or down based on workload demands. The platform can automatically handle workload distribution across multiple instances, ensuring efficient resource utilization and adapting to increasing traffic or compute demands.

[0163] High Availability: Container management platforms typically provide built-in mechanisms to ensure the high availability of language models. By deploying language models within a container management platform, replica sets can be configured, or multiple instances of the model can be deployed on different nodes within the cluster. This redundancy ensures that the language model remains available even if one or more instances or nodes fail or require downtime for maintenance.

[0164] Resource Management: The container management platform allows you to define resource limits and requests for individual containers or pods. By specifying the required CPU and / or memory resources for each language model instance, the container management platform can efficiently allocate and / or manage available resources within the cluster. This ensures that language model instances have the resources needed for optimal execution without impacting other applications running on the same cluster.

[0165] Load balancing: Container management platforms can provide load balancing capabilities to distribute incoming requests across multiple instances of the language model. This helps to evenly distribute the compute load, preventing any single instance from becoming overwhelmed. Load balancing enhances the performance and responsiveness of the language model, enabling it to effectively handle increased user traffic.

[0166] Rolling Updates and Rollbacks: The container management platform supports rolling updates, allowing for updates to the language model while minimizing downtime. By gradually rolling out new versions of model instances, the platform enables seamless updates without interrupting ongoing requests or user interactions. Furthermore, if any issues arise during the update process, the platform can easily roll back to a previous working version, ensuring the availability and stability of the language model.

[0167] Resource monitoring and auto-scaling: Container management platforms offer extensive monitoring and / or metric collection capabilities. By leveraging monitoring tools and integrating them with metric collectors such as Prometheus, valuable insights into the resource utilization, performance, and / or health of language model instances can be gathered. Based on these metrics, container management platforms can automatically increase or decrease the number of instances through horizontal Pod auto-scaling, ensuring optimal resource utilization and maintaining the required performance levels.

[0168] Infrastructure Portability: Container management platforms typically offer a high level of infrastructure portability. Regardless of whether the language model is running on-premises, in the public cloud, or in a hybrid environment, the container management platform provides a consistent framework for managing and / or deploying the model. This portability allows for the flexible selection of the infrastructure best suited to the needs and simplifies the migration or deployment of the language model across different environments.

[0169] According to another aspect of this disclosure, a method for training a language model is provided. This method can train the language model for use in computer-aided troubleshooting of a packaging line configured for production packaging, particularly for any method disclosed herein. This method (i.e., the training method) may include the step of providing a language model (i.e., an initial untrained model). The initial model may contain any features of the language model disclosed herein, such as a Transformer architecture. The method may include the step of training the language model based on training data. The training data may include one or more of the following: at least one operating manual associated with the packaging line; process data collected from the packaging line; stored information indicating previously encountered problems and corresponding previously generated instructions of one or more actions to be performed to mitigate the problem; and / or more generally, any training data disclosed herein. Training may be performed using self-supervised training, as described in other parts of this disclosure.

[0170] All the methods disclosed in this article can be implemented, in part or in whole, by a computer.

[0171] According to another aspect of this disclosure, a language model is provided that has been trained for any of the methods disclosed herein, particularly using the training methods described above.

[0172] According to another aspect of this disclosure, a data processing apparatus is provided, comprising means for performing any of the methods disclosed herein. According to another aspect of this disclosure, a packaging line including such a data processing apparatus is provided.

[0173] According to another aspect of this disclosure, a computer program or a computer-readable medium storing a computer program is provided, the computer program including instructions that, when executed by a computer, cause the computer to perform any of the methods disclosed herein.

[0174] Other objects, features, aspects and advantages of the invention will become apparent from the following detailed description and the accompanying drawings. Attached Figure Description

[0175] The following figures will provide a better understanding of this disclosure:

[0176] Figure 1 This is a schematic block diagram of the packaging line according to the implementation scheme of this disclosure;

[0177] Figure 2 A conceptual diagram of training data that can be used to train a language model according to an implementation scheme of this disclosure;

[0178] Figure 3 Example pages for a manual that can be used as language model training data according to the implementation of this disclosure;

[0179] Figure 4 A conceptual diagram of the input and output of a language model trained according to an implementation scheme of this disclosure;

[0180] Figure 5 A flowchart illustrating a method for training a language model according to an implementation scheme of this disclosure;

[0181] Figure 6 This is a flowchart illustrating a method for computer-aided troubleshooting in a packaging line, based on an embodiment of this disclosure. Detailed Implementation

[0182] The representative embodiments shown in the accompanying drawings will be explained below. It should be understood that the embodiments shown and the following description are not intended to limit the embodiments to a single preferred embodiment.

[0183] Figure 1This is a schematic block diagram of packaging line 1 in an embodiment of this disclosure. In the illustrated embodiment, packaging line 1 includes one or more packaging components or machines, with a non-limiting example in the figure being a filling machine 2 configured to produce packages from tubes of packaging material containing pourable (food) products. Food refers to any substance that humans and animals can ingest, eat, and / or drink, including liquids, semi-liquids, viscous substances, dry powders, powders, and solid foods, beverages, and water. It should be understood that although only filling machine 2 is shown in the figure, packaging line 1 may include multiple machines and equipment configured according to the general knowledge of those skilled in the art for transporting and producing packages, such as different types of filling machines (which produce packages from packaging material blanks or tubes), downstream processing equipment, or warehouses for transporting and / or storing rolls of packaging material and / or packages. Furthermore, the solutions of this disclosure are applicable to any packaging plant, including any processing machinery, and to any type of pourable food. Packaging line 1 can be configured to produce packaging at high speed, for example at a rate of at least 8,000 packages per hour, preferably at least 12,000 packages per hour, and even more preferably at least 24,000 packages per hour.

[0184] Packaging line 1 includes a data processing device 4 configured to perform computer-aided troubleshooting. The data processing device 4 includes a processing unit 40 configured to execute computer-aided troubleshooting methods. The data processing device 4 and / or the processing unit 40 may be located locally or remotely (on a remote server or in the "cloud") on packaging line 1. Therefore, packaging line 1 may include remote computing capabilities.

[0185] Packaging line 1 may include multiple sensors 5 arranged along its length. Although only one sensor 5 is depicted in the figure, it should be understood that multiple sensors 5 may be arranged not only on the filling machine 2, but also along one or more machines or devices along packaging line 1. The multiple sensors 5 may be configured to collect process data PD from packaging line 1.

[0186] Packaging line 1 may include actuators configured to perform operations according to a control signal C, such as the control signal C indicating the setup of packaging line 1 and / or events within packaging line 1. By way of example only, actuators may include sealing devices, rollers, final folding devices, etc., within filling machine 2. It should be understood that actuators may be arranged along packaging line 1.

[0187] The packaging line 1 may include a control unit 6 configured to control the operation of the packaging line 1. The control unit 6 may be directly or indirectly connected to a plurality of sensors 5 and / or a plurality of actuators. The control unit 6 may be configured to control actuators and / or a plurality of sensors 5. A data processing device 4 may be (e.g., directly or indirectly) connected to the control unit 6. Optionally, the data processing device 4 may be (e.g., directly or indirectly) connected to a plurality of sensors 5 to directly receive sensor signals from them. In some embodiments, the control unit 6 and the data processing device 4 may be integrated into a single device.

[0188] Processing unit 40 may be configured to apply language model 42, such as a trained language model disclosed elsewhere herein. Processing unit 40 also includes a database 44 configured to store information available to language model 42. It should be understood that database 44 may also be located elsewhere outside of processing unit 40. Figure 1 A user interface 10 is also shown, which is configured to allow users to interact with the language model 42. For this purpose, the user interface 10 may include suitable input and output devices.

[0189] Figure 2 A conceptual block diagram 200 illustrates training data that can be used to train language model 42 according to an embodiment of this disclosure. It can be seen that language model 42 can be used using one or more operation manuals 202 associated with packaging line 1, process data PD collected from packaging line 1, and / or information stored in database 44 indicating previously encountered problems and corresponding previously generated instructions for actions to be performed to alleviate those problems. In a preferred embodiment, language model 42 is trained using self-supervised training as described elsewhere herein. Language model 42 may be based on the Transformer architecture described elsewhere herein.

[0190] Figure 3 An exemplary page of the instruction manual 202 is shown, which, in one implementation, can be used as training data for the language model 42.

[0191] Figure 4 A conceptual block diagram 400 illustrates the inputs and outputs of a trained language model 42 during runtime (also known as inference). It can be seen that the language model 42 receives a problem statement 402 in natural language from the user as input, indicating a problem that arises during the operation of packaging line 1. The problem statement 402 can be received via the user interface 10 (particularly via its input device). The language model 42 generates one or more instructions 404 as output, instructing the user on actions to be performed on packaging line 1 to alleviate the problem. One or more instructions 404 can be presented to the user via the user interface 10 (particularly via its output device). Figure 4 It is also shown that the language model 42 can take process data PD as input and generate one or more instructions 404 based on it. The figure also shows that the language model 42 can generate control signals C as output.

[0192] Figure 5 A method 500 for training a language model 42 is illustrated. In step 502, an initial, untrained language model 42 is provided. In step 504, the language model 42 is trained, preferably using a method such as combining... Figure 2 The self-supervised training using training data.

[0193] Figure 6 A method 600 for computer-aided troubleshooting of packaging line 1 is illustrated. In step 602, a problem statement 402 is received. The problem statement 402, received from the user in natural language, indicates a problem occurring during the operation of packaging line 1. In step 604, a language model 42 generates one or more instructions 404. The one or more instructions 404 instruct the user to perform an action on packaging line 1 to alleviate the problem. In step 606, the one or more instructions 404 are output to the user in natural language.

[0194] Although some aspects are described in the context of an apparatus, they clearly also represent a description of the corresponding method, where modules or devices correspond to method steps or features of method steps. Similarly, aspects described in the context of method steps also represent a description of corresponding blocks, items, or features of the corresponding apparatus.

[0195] Implementations of this disclosure can be carried out on a computer system. This computer system can be a local computer device (e.g., a personal computer, laptop, tablet, or mobile phone) having one or more processors and one or more storage devices, or a distributed computer system (e.g., a cloud computing system having one or more processors and one or more storage devices distributed in different locations (e.g., on local clients and / or one or more remote server clusters and / or data centers)). The computer system can include any circuitry or combination of circuits. In one embodiment, the computer system can include one or more processors, which can be of any type. As used herein, "processor" can refer to any type of computing circuitry, such as, but not limited to, microprocessors, microcontrollers, complex instruction set computing (CISC) microprocessors, reduced instruction set computing (RISC) microprocessors, very long instruction word (VLIW) microprocessors, graphics processing units (GPUs), digital signal processors (DSPs), multi-core processors, field-programmable gate arrays (FPGAs), or any other type of processor or processing circuitry. Other types of circuitry that may be included in the computer system can be custom circuitry, application-specific integrated circuits (ASICs), etc., such as one or more circuits (e.g., communication circuitry) for wireless devices such as mobile phones, tablets, laptops, two-way radios, and similar electronic systems. A computer system may include: one or more storage devices, which may include one or more application-specific storage elements, such as main memory in the form of random access memory (RAM); one or more hard disk drives; and / or one or more drives for processing removable media (such as optical discs (CDs), flash memory cards, digital video discs (DVDs), etc.). The computer system may also include a display device, one or more speakers, and a keyboard and / or controller, wherein the keyboard and / or controller may include a mouse, trackball, touchscreen, voice recognition device, or any other device that allows system users to input information into and receive information from the computer system.

[0196] Some or all of the method steps can be performed by (or using) hardware devices (such as processors, microprocessors, programmable computers, or electronic circuits). In some embodiments, one or more of the most important method steps can be performed by such devices.

[0197] Depending on the specific implementation requirements, embodiments of this disclosure can be implemented in hardware or software. This implementation can be executed using non-transitory storage media (e.g., digital storage media such as floppy disks, DVDs, Blu-ray discs, CDs, ROMs, PROMs, EPROMs, EEPROMs, or flash memory) that store electronically readable control signals that can (or are capable of) cooperating with a programmable computer system to execute the corresponding methods. Therefore, the digital storage media can be computer-readable.

[0198] Some embodiments of this disclosure include a data carrier having electronically readable control signals that can cooperate with a programmable computer system to perform one of the methods described herein.

[0199] Typically, embodiments of this disclosure can be implemented as a computer program product containing program code that, when run on a computer, operates to perform one of the methods. For example, the program code can be stored on a machine-readable medium.

[0200] Other implementations include a computer program for performing one of the methods described herein, the computer program being stored on a machine-readable medium.

[0201] Therefore, in other words, an embodiment of this disclosure is a computer program having program code for performing one of the methods described herein when run on a computer.

[0202] Therefore, another embodiment of this disclosure is a storage medium (or data carrier, or computer-readable medium) having a computer program stored thereon for performing one of the methods described herein when executed by a processor. Data carriers, digital storage media, or recording media are generally tangible and / or non-transitory. Another embodiment of this disclosure is an apparatus as described herein, comprising a processor and a storage medium.

[0203] Therefore, another embodiment of this disclosure represents a data stream or signal sequence for performing one of the methods described herein. This data stream or signal sequence may, for example, be configured to be transmitted via a data communication connection (e.g., via the Internet).

[0204] Another implementation includes a processing mechanism, such as a computer or programmable logic device, configured or adapted to perform one of the methods described herein.

[0205] Another embodiment includes a computer on which a computer program is installed for performing one of the methods described herein.

[0206] Another embodiment of this disclosure includes an apparatus or system configured to transmit a computer program (e.g., electronically or optically) for performing one of the methods described herein to a receiver. The receiver may be, for example, a computer, a mobile device, a memory device, etc. The apparatus or system may, for example, include a file server for transmitting the computer program to the receiver.

[0207] In some embodiments, a programmable logic device (e.g., a field-programmable gate array) may be used to perform some or all of the functions of the methods described herein. In some embodiments, the field-programmable gate array may cooperate with a microprocessor to perform one of the methods described herein. Generally, these methods are preferably performed by any hardware device.

[0208] As can be seen from the above description, although various embodiments of the invention have been described and illustrated, the invention is not limited thereto, but may be embodied in other ways within the scope of the subject matter defined in the following claims.

Claims

1. A method (600) for computer-aided troubleshooting of a packaging line (1) configured to produce packaging, the method (600) comprising: The input device receives (602) a natural language problem statement (402) from the user, which indicates a problem that has occurred during the operation of the packaging line (1); One or more instructions (404) are generated (604) at least in part based on the problem statement (402) by a trained language model (42), the one or more instructions (404) indicating an action to be performed by the user on the packaging line (1) to alleviate the problem; as well as The one or more instructions (404) are output to the user in natural language (606) via an output device.

2. The method according to claim 1, wherein, The language model (42) has been trained using at least one operation manual (202) associated with the packaging line (1).

3. The method according to claim 1 or 2, wherein, The language model (42) has been trained using process data (PD) collected from the packaging line (1).

4. The method according to any one of the preceding claims, wherein, The language model (42) has been trained using stored information (44) indicating previously encountered problems and corresponding previously generated one or more instructions, which indicate actions to be performed to alleviate the problems.

5. The method according to any one of the preceding claims, wherein, The step of generating (604) one or more instructions (404) is also based at least in part on process data (PD) collected from the packaging line (1).

6. The method according to any one of the preceding claims, wherein, The step of generating (604) one or more instructions (404) includes generating a plurality of instructions (404), and wherein the step of outputting (606) includes outputting the plurality of instructions (404) as a sequence of operations to be performed by a user.

7. The method according to any one of the preceding claims, wherein, The steps of generating one or more instructions (604) further include: In response to receiving the question statement (402), one or more dialogue cycles are performed, including: The trained language model (42) generates a specific request in natural language to prompt the user to provide additional information related to the question; The output device outputs the specific request for the natural language to the user. The user receives additional information related to the question in natural language via the input device. One or more candidate instructions are generated by the trained language model (42), which indicate the operation to be performed by the user on the packaging line (1) to alleviate the problem; Specifically, one or more dialogue cycles are executed until one or more candidate instructions reach a predetermined confidence score; and The output step (606) includes outputting one or more candidate instructions to the user in natural language.

8. The method according to any one of the preceding claims, further comprising: At least one control signal (C) is generated based at least in part on one or more instructions (404), the control signal (C) being configured to control and / or modify the operation of the packaging line (1) to alleviate the problem.

9. The method according to any one of the preceding claims, wherein, The language model (42) has been trained using self-supervised training.

10. The method according to any one of the preceding claims, wherein, The language model (42) includes neural networks, particularly deep neural networks and / or Transformer architectures.

11. The method according to any one of the preceding claims, wherein, The language model (42) is deployed on a container management platform.

12. A method (500) for training a language model (42), said language model (42) for computer-aided troubleshooting of a packaging line (1) configured to produce packaging, particularly for the method (500) according to any one of claims 1-11, wherein, The method (500) includes: Provides (502) language models (42); The language model (42) is trained (504) based on training data, wherein the training data includes one or more of the following: At least one operation manual (202) related to the packaging line (1); Process data (PD) collected from the packaging line (1); and / or The stored information (44) indicates a previously encountered problem and one or more instructions of the corresponding previously generated instructions to be executed to alleviate the problem; Optionally, the training (504) is performed using self-supervised training.

13. A language model (42) trained for use in the method (600) of any one of claims 1-11, particularly trained using the training method (500) of claim 12.

14. A data processing apparatus (4) or a packaging line (1) including a data processing apparatus (4), the data processing apparatus (4) including a mechanism for performing the method (500, 600) according to any one of claims 1-12.

15. A computer program or a computer-readable medium storing a computer program, the computer program comprising instructions that, when a computer executes the program, cause the computer to perform the method (500, 600) according to any one of claims 1-12.

Images