Packaging bag intelligent production management method and system

Abstract

The application discloses a kind of packing bag wisdom production management method and system, it is related to production management field, when AOI camera detects original defect signal, system immediately generates the defect event containing detailed production context, breaks the traditional information fragmentation situation.Subsequently, with defect timestamp as center, system initiates query to time series database, obtains PLC key parameter and MES material batch information, forms comprehensive defect context fingerprint, solves data island problem.This fingerprint is compared with digital twin baseline, identifies abnormal vector, reveals potential process deviation.Finally, abnormal vector is input into pre-trained machine learning model, combined with knowledge base, intelligently infers defect reason and provides accurate recommended action, and is displayed in real time on the screen in central control room, thereby significantly shortening root cause positioning time, realizes the intelligentization and lean of production management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of production management, and more specifically, to a smart production management method and system for packaging bags. BACKGROUND

[0002] In modern industrial production, the production management of packaging bags is facing increasingly complex challenges. With the increasing demand for product quality and production efficiency in the market, the traditional production management mode has been difficult to meet the needs. Especially in the high-speed printing link, if defects occur, if the cause cannot be traced and corrective measures are taken in time and accurately, a large amount of waste products will be produced, which will seriously affect the production cost and delivery cycle. Therefore, it is of great significance to build an efficient and intelligent packaging bag smart production management scheme to realize real-time monitoring of the production process, rapid positioning of defects and root cause analysis, which is of great significance to improve the competitiveness of enterprises.

[0003] However, in the prior art, although automatic equipment and information systems may have been introduced in each production link, these systems often operate independently, forming a data island. For example, an automatic optical inspection (AOI) system can find product defects and generate defect pictures, but its information is usually limited to the defect itself; a manufacturing execution system (MES) records production orders, material batches, etc. However, it cannot be directly associated with device operating parameters; a programmable logic controller (PLC) collects device operating data such as pressure, speed, temperature, etc. in real time, but these data are usually only for reference by device operators, and it is difficult to effectively integrate with other systems. When defects such as light-colored ink smearing occur on the production line, the AOI operator, the field foreman, the MES system and the PLC system each have part of the information, but these key information (such as AOI defects, PLC parameter fluctuations, MES material replacement records, and foreman operations) is fragmented, and no one can instantly obtain global information. This leads to inefficient defect-cause tracing and delayed root cause positioning. For example, the foreman may have adjusted the doctor blade pressure based on experience, but the real root cause is that the insufficient drying temperature caused the ink to be not dry, and during this period, hundreds of meters of waste products have been produced. This situation of information fragmentation and data island makes it difficult for enterprises to respond quickly and make accurate decisions when facing production abnormalities, which seriously restricts the further improvement of production efficiency and product quality.

[0004] Therefore, there is an urgent need for a packaging bag smart production management scheme that can break down data islands, integrate multi-source information, and realize real-time defect tracing and intelligent root cause positioning. SUMMARY

[0005] Based on the defects existing in the prior art, according to an aspect of the present application, a packaging bag intelligent production management method is provided, which comprises: in response to an original defect signal from an AOI camera, generating a defect event based on the original defect signal; taking the timestamp of the defect event as the center, initiating a range query to a time series database to obtain a defect context fingerprint; comparing the defect context fingerprint with a digital twin baseline to obtain an abnormal vector; inputting the abnormal vector into a pre-trained multi-label classification machine learning model to obtain a cause probability; inputting the cause probability into a knowledge base to obtain a recommended action; displaying the defect event, the cause probability and the recommended action on a screen in a control room.

[0006] According to another aspect of the present application, a packaging bag intelligent production management system is provided, which comprises: a defect event generation module for generating a defect event based on an original defect signal from an AOI camera in response to the original defect signal; a defect context fingerprint acquisition module for taking the timestamp of the defect event as the center, initiating a range query to a time series database to obtain a defect context fingerprint; an abnormal vector generation module for comparing the defect context fingerprint with a digital twin baseline to obtain an abnormal vector; a cause probability generation module for inputting the abnormal vector into a pre-trained multi-label classification machine learning model to obtain a cause probability; a recommended action generation module for inputting the cause probability into a knowledge base to obtain a recommended action; and a display module for displaying the defect event, the cause probability and the recommended action on a screen in a control room.

[0007] Compared with the prior art, the packaging bag intelligent production management method and system provided by the present application breaks the traditional information fragmentation when the AOI camera detects an original defect signal, and the system immediately generates a defect event containing detailed production context. Subsequently, taking the defect timestamp as the center, the system initiates a query to a time series database to obtain PLC key parameters and MES material batch information, forming a comprehensive defect context fingerprint, and solving the data island problem. Comparing the fingerprint with a digital twin baseline, the abnormal vector is identified, and the potential process deviation is revealed. Finally, the abnormal vector is input into a pre-trained machine learning model, combined with a knowledge base, to intelligently infer the defect cause and provide accurate recommended actions, and to display in real time on a screen in a control room, so as to improve the defect-cause traceability from inefficient manual experience judgment to efficient intelligent analysis, significantly shorten the root cause positioning time, reduce the scrap rate, and realize the intelligentization and leanization of production management. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description thereof taken in conjunction with the accompanying drawings, in which: The accompanying drawings provide exemplary embodiments of the application and serve to illustrate the principles of the present application. The drawings are not intended to limit the scope of the application in any way but rather serve to explain the principles of the application. In the drawings:

[0009] Figure 1 Flow chart of the packaging bag intelligent production management method according to the embodiment of the present application.

[0010] Figure 2 Data flow diagram of the packaging bag intelligent production management method according to the embodiment of the present application.

[0011] Figure 3 Flow chart of step S1 in the packaging bag intelligent production management method according to the embodiment of the present application.

[0012] Figure 4 Flow chart of step S3 in the packaging bag intelligent production management method according to the embodiment of the present application.

[0013] Figure 5 Block diagram of the packaging bag intelligent production management system according to the embodiment of the present application. DETAILED DESCRIPTION

[0014] Embodiments of the present disclosure will be described in more detail by referring to the drawings. While the present disclosure is shown and described in connection with certain embodiments thereof, it is to be understood that the present disclosure is capable of further minor variations and modifications and can be practiced by other than the embodiments described herein. Accordingly, it is submitted that that the present disclosure shall not be limited to the embodiments set forth herein for purposes of elucidation.

[0015] In view of the technical defects exposed by the above background art, the present application proposes a packaging bag intelligent production management method. Figure 1 Flow chart of the packaging bag intelligent production management method according to the embodiment of the present application. Figure 2 Data flow diagram of the packaging bag intelligent production management method according to the embodiment of the present application. As Figure 1 and Figure 2As shown, the package bag intelligent production management method according to the embodiment of the application comprises: S1, in response to an original defect signal from an AOI camera, generating a defect event based on the original defect signal; S2, taking the timestamp of the defect event as the center, initiating a range query to a time series database to obtain a defect context fingerprint; S3, comparing the defect context fingerprint with a digital twin baseline to obtain an anomaly vector; S4, inputting the anomaly vector into a pre-trained multi-label classification machine learning model to obtain a cause probability; S5, inputting the cause probability into a knowledge base to obtain a recommended action; and S6, displaying the defect event, the cause probability and the recommended action on a screen in a central control room.

[0016] In step S1, in response to an original defect signal from an AOI camera, a defect event is generated based on the original defect signal. It should be understood that although the traditional AOI system can detect defects, the original signal output by the traditional AOI system often only contains image data and simple defect position information, and lacks association with key context information such as production process, equipment state, material batch, etc. This information fragmentation leads to the data silo problem mentioned in the background art, making it difficult to quickly and accurately trace the root cause of the defect after the defect occurs. Therefore, by converting the original defect signal into a defect event containing rich context information, the application can preliminarily integrate the data scattered in different systems (such as AOI, MES, PLC), providing a unified and complete data basis for subsequent analysis. This enables the management scheme to shift from passive post-detection to active real-time analysis and prevention, significantly improving defect tracing efficiency and the accuracy of root cause positioning, thereby effectively reducing the scrap rate and optimizing production management.

[0017] Specifically, in one exemplary embodiment of the application, Figure 3 The flow chart of step S1 in the package bag intelligent production management method according to the embodiment of the application. As shown in FIG. 2, the method comprises the following steps: Figure 3As shown, in step S1, in response to the original defect signal from the AOI camera, a defect event is generated based on the original defect signal, including: S11, extracting the camera ID, timestamp, original image data, defect pixel coordinates and pulse count from the original defect signal; S12, generating a unique event ID; S13, based on the camera ID, obtaining the device ID by querying the table; S14, based on the camera ID, initiating a synchronous API call to the MES system to obtain the production order signal, the production order signal including the current work order number, current product code and current production batch; S15, based on the pulse count, calculating the accurate meter mark; S16, performing preliminary defect classification and severity assessment on the original image data to obtain a preliminary classification result and a severity score; S17, compressing and cropping the original image data to obtain an image accessible URL; S18, assembling the unique event ID, the device ID, the timestamp, the current work order number, the current product code, the current production batch, the accurate meter mark, the classification result, the severity score and the image accessible URL to obtain the defect event.

[0018] It can be understood that these information is the starting point of building a complete defect event: the camera ID is used to locate the specific device where the defect occurs; the timestamp records the exact time when the defect occurs, which is the time reference for all subsequent data correlation; the original image data is the visual evidence of the defect, which is used for subsequent classification and evaluation; the defect pixel coordinates indicate the specific location of the defect in the image, which helps accurate positioning; and the pulse count is the key to calculating the accurate meter mark of the defect on the production line, ensuring the physical accuracy of the defect position.

[0019] The implementation details of step S11 can be further described as follows: When the AOI camera detects a defect on the surface of the packaging bag on the production line, it will immediately generate an original defect signal. This signal is transmitted to the data processing module through industrial Ethernet or high-speed data bus in a specific data packet format. The data processing module has a pre-set parser for the AOI camera data protocol, which can identify and extract each field in the data packet. First, the parser reads the camera ID from the header or a specific field of the data packet. For example, if the data packet contains a field named Camera_ID with the value AOI-001, it is extracted as the camera ID. This camera ID is a unique identifier set in advance when the AOI camera is configured, used to distinguish different detection devices. Second, the parser extracts the timestamp from the data packet. The AOI camera synchronously records the current time information at the moment of detecting the defect, and embeds it into the data packet of the original defect signal. For example, the data packet may contain a Timestamp field with the value 2024-7-27 10:35:12.345, which is extracted as the timestamp of the defect occurrence. Third, the parser extracts the original image data from the data packet. When the AOI camera detects a defect, it captures an image containing the defect area and encapsulates it in binary stream form in the original defect signal. The parser will extract this part of binary data completely according to the pre-defined image data start position and length in the data packet, forming the original image data. For example, it may be a JPEG or PNG format image data stream. At the same time, the parser also parses the defect pixel coordinates from the original defect signal. The AOI camera will identify the specific position of the defect area in the image during internal processing, and convert it into pixel coordinates in the image coordinate system, such as representing the center point or the top-left and bottom-right coordinates of the bounding box of the defect in the form of (x, y). These coordinate information will be included in the original defect signal as a specific field, and the parser will accurately extract these pixel coordinates, for example, (120, 350) in the Defect_Coord field. Finally, the parser extracts the pulse count. On the packaging bag production line, an encoder or pulse sensor is installed to accurately measure the number of meters or products produced. When the AOI camera detects a defect, it synchronously acquires the current reading of the encoder or pulse sensor and includes it as the pulse count in the original defect signal. For example, the data packet may contain a Pulse_Count field with the value 56789.

[0020] It can be understood that in a high-speed production environment, defects can occur frequently, and different defects can have similar characteristics or occur at similar time points. Without a globally unique identifier, it will be difficult to accurately distinguish and track each specific defect event, thereby affecting subsequent data correlation, analysis and decision-making. By assigning a globally unique ID to each defect event, it can be ensured that each defect event can be accurately identified and managed even in a distributed system or after a long time, providing a stable reference basis for subsequent defect context fingerprint construction, anomaly vector comparison, cause probability inference and recommended actions.

[0021] The implementation details of step S12 can be further described as follows: the unique event ID is the anchor point for all subsequent data associated with the defect event, ensuring that the traceability and management of a specific defect in the entire intelligent production management process have clear directionality, avoiding data confusion and conflicts. This process assigns a globally unique identifier to each defect instance currently detected to ensure its traceability and independence in the entire production management process. Specifically, the data processing module will call a special ID generation service or library function to complete this task. The service or library function is implemented based on the Universal Unique Identifier (UUID) standard. UUID is a standard defined by the Open Software Foundation (OSF) that can guarantee uniqueness in time and space, even without central coordination, to generate non-repeating identifiers. In actual operation, when receiving the data output by step S11 (including camera ID, timestamp, original image, defect pixel coordinates and pulse count), the system will trigger the ID generation service. The service integrates the current timestamp, device MAC address or IP address, random number and counter, and uses standard algorithms such as UUID version 1 or version 4 to calculate a 128-bit unique identifier. This identifier serves as the unique event ID for this defect event.

[0022] It is easy to understand that the AOI camera, as an independent detection unit, is deployed on the production line, and its own ID only identifies the detection device, but cannot directly reflect the specific production equipment it monitors. In order to correlate and analyze defect information with specific production equipment and its operating status, historical data, etc., a mapping relationship between the camera ID and the production equipment ID needs to be established. By querying the pre-set mapping table, the production equipment ID corresponding to the defect occurrence can be quickly and accurately obtained, thereby providing accurate equipment attribution information for subsequent analysis, effectively breaking down the data silos at the equipment level.

[0023] The implementation details of step S13 can be further described as follows: After successfully generating a unique event ID, the data processing module will use the camera ID as input to perform a query operation to obtain the corresponding device ID. This process relies on a pre-constructed and maintained camera-device mapping query table. In particular, this query table is a structured data storage, existing in the form of a database table or a hash table in memory. The table contains at least two columns of key information: one column stores the camera ID, and the other column stores the device ID corresponding to the camera ID. During the production line deployment and configuration phase, detailed device asset management is performed, binding each AOI camera to the specific production device it monitors, and recording this binding relationship in the query table. For example, if the AOI-001 camera is installed on the 2nd printing unit, there will be a record in the query table mapping AOI-001 to PRINTING-UNIT-002. When the data processing module receives the camera ID, such as AOI-001, it will initiate a query request to the camera-device mapping query table. The query request will use the camera ID as the query key to find the matching record in the table. Once the matching record is found, the data processing module will extract the corresponding device ID, such as PRINTING-UNIT-002, as the corresponding device ID.

[0024] Correspondingly, the occurrence of defects is often closely related to the current production task, the type of product being produced, and the batch of materials used. Traditional defect detection only stays at the image level and cannot directly obtain these production context information. By interacting with the MES system in real time, the work order number, product code, and production batch being executed at the time of defect occurrence can be obtained, thereby connecting the defect information with the specific production task and material traceability chain, providing important production background information for subsequent defect cause analysis, and effectively solving the problem of fragmented production information.

[0025] The implementation details of step S14 can be further described as follows: After obtaining the device ID corresponding to the camera ID, the data processing module will use the device ID as a parameter to initiate a synchronous API (Application Programming Interface) call to the MES (Manufacturing Execution System). This API call aims to query the production task information currently being executed by the device in the MES system in real time. The MES system provides a standard API interface that allows external applications to query the production status. The data processing module will construct a request that conforms to the MES system API specification, containing the obtained device ID and timestamp. When the MES system receives this API request, it will query its internal production plan and execution records based on the incoming device ID and timestamp. The MES system maintains the real-time status of each production device, including the current work order being executed, the product being produced, and the material batch being used. The MES system will generate a production order signal as a response to the API call based on the queried information. This production order signal is a structured data packet returned in JSON or XML format. It contains at least the following key information: the current work order number to identify the unique number of the production task currently being executed by the device. The current product code to identify the unique code of the product currently being produced by the device. The current production batch to identify the material batch or production batch number to which the product currently being produced belongs. After initiating the synchronous API call, the data processing module waits for the response from the MES system. Once the response is received, it parses the response data packet and extracts the current work order number, current product code, and current production batch from it.

[0026] It can be understood that the defects detected by the AOI camera are usually presented in pixel coordinates, while defect tracing and repair in the production site often need to be positioned in actual physical length (meter scale). Traditional defect positioning may rely on manual visual inspection or rough estimation, which is inefficient and lacks accuracy. By converting pulse counts to precise meter scale, abstract pixel coordinates can be mapped to actual production length, enabling precise physical positioning of defects on the production line, facilitating operators to quickly find defect locations and process them, effectively improving defect tracing efficiency and accuracy.

[0027] The implementation details of step S15 can be further described as follows: After obtaining the production order signal, the data processing module will use the pulse count as input to perform the calculation of the accurate meter mark. This process relies on the pre-set pulse equivalent parameter. The pulse equivalent refers to the number of pulses of the pulse sensor or encoder corresponding to a unit length, for example, 1 meter. This parameter is determined through accurate measurement and calibration during the installation and commissioning of the equipment, and is associated with the equipment ID, stored in the configuration parameters of the data processing module or in the queryable equipment parameter database. For example, if the encoder installed on the production line generates 1000 pulses per revolution, and the roller connected to the encoder has a circumference of 0.5 meters, then the number of pulses per meter is 1000 / 0.5 = 2000 pulses / meter. The formula for the accurate meter mark is: accurate meter mark = pulse count / pulse equivalent. In implementation, the data processing module obtains the current pulse count, for example, 56789. At the same time, according to the equipment ID, the pulse equivalent of the equipment is queried from the pre-set equipment parameters, such as 2000. Then, the formula is calculated to obtain the accurate meter mark as 56789 / 2000 = 28.3945 meters. The calculated accurate meter mark is the accurate physical position of the product on the production line when the defect occurs.

[0028] It is worth mentioning that although the AOI camera can capture defect images, the original image data itself cannot directly provide the type (such as light color ink smearing, scratches, foreign matter, etc.) and severity of the defect. Manual visual inspection is inefficient and easily affected by subjective factors. By automatically classifying and assessing the severity of the defect image, the nature of the defect can be quickly identified, and the severity can be prioritized according to the severity, thereby guiding subsequent production intervention and quality control, effectively improving the efficiency and accuracy of defect handling, and avoiding losses caused by information lag or judgment errors.

[0029] The implementation details of step S16 can be further described as follows: this process mainly relies on a pre-trained deep learning model, such as a convolutional neural network (CNN) model. The deep learning model adopts a classic CNN architecture such as ResNet, VGG or EfficientNet, and is optimized for the characteristics of packaging bag defect images. The input layer of the model is designed to receive raw image data, the middle layer contains multiple convolutional layers, pooling layers and activation function layers, which are used to extract features from the image, and finally the result is output through a fully connected layer and a Softmax activation function. In particular, at the output end of the model, two branches are designed, one for outputting the probability distribution of each defect type as a preliminary classification result, and the other for outputting a continuous value between 0 and 1 representing the severity score of the defect. The model acquires defect recognition ability through the training process. First, a large number of packaging bag defect images are collected and labeled by humans for defect type and severity level; then these labeled data are input into the CNN model, and the backpropagation algorithm and optimizer are used to continuously adjust the model parameters to minimize the prediction error; after multiple rounds of iterative training, the model gradually learns the visual features of each defect type and its severity pattern, and finally acquires the ability to accurately classify and evaluate new defect images. In specific implementation, when the data processing module receives raw image data, it inputs the image data into the trained deep learning model. The model performs forward propagation calculation on the image and outputs a vector containing the probability of each defect category, as well as a numerical value representing the severity level. For example, for a raw image data of a light-colored ink smearing, the model outputs: a preliminary classification result of the defect, i.e. a probability distribution, such as: {light-colored ink smearing: 0.95, scratch: 0.02, foreign matter: 0.01, others: 0.02}. The data processing module selects the category with the highest probability as the preliminary classification result, i.e. light-colored ink smearing. The severity score is a floating-point number between 0 and 1, for example: 0.85.

[0030] Accordingly, since the raw image data captured by the AOI camera usually has high resolution and large file size, direct storage and transmission will occupy a large amount of storage space and network bandwidth, affecting system performance. At the same time, when displaying defect images on the user interface, usually only a thumbnail or a cropped local image is needed for quick preview. By compressing and cropping the image, the data volume can be significantly reduced, improving storage and transmission efficiency, and by generating an accessible URL, users or other modules can view the defect image anytime and anywhere, effectively improving the usability and response speed of the system.

[0031] The implementation details of step S17 can be further described as follows: Specifically, in an example embodiment of the present application, step S17, compressing and cropping the original image data to obtain an image accessible URL, comprises: S171, compressing and cropping the original image data to obtain a thumbnail image; S172, uploading the thumbnail image to an object storage service to obtain the image accessible URL.

[0032] First, step S171 is performed. The data processing module receives original image data, such as a PNG image file with a resolution of 2048x1536 pixels, and defect pixel coordinates, such as a bounding box [(100, 300), (150, 400)] representing a defect area. The data processing module extracts the local area containing the defect from the original image according to the defect pixel coordinates. In order to provide sufficient context information, the cropped area will be slightly larger than the defect itself, for example, extending 20-50 pixels outward from the center of the defect bounding box. If the original image data does not provide defect pixel coordinates, the center region or the entire image can be cropped. Then, the cropped image is compressed to reduce file size. Common image compression algorithms include JPEG, WebP, etc. The compression parameters (such as JPEG quality factor) can be preset, for example, setting the JPEG quality factor to 75 to achieve a higher compression ratio while ensuring visual quality. At the same time, the resolution of the image is also adjusted to a size suitable for preview, for example, scaling the cropped image to 200x200 pixels or 400x400 pixels to generate a thumbnail image. Then, step S172 is performed. After generating the thumbnail image, the data processing module uploads the thumbnail file to an object storage service. The object storage service is a highly available, highly scalable, and highly reliable cloud storage service, such as Amazon S3, Aliyun OSS, or MinIO, etc. The upload process is completed by calling the API interface of the object storage service. After successful upload, the object storage service returns an image accessible URL. This URL is a public or access-allowed link that allows direct access and download of the thumbnail image through HTTP / HTTPS protocol.

[0033] It can be understood that the above steps respectively obtain various independent information about the defect from different sources through different processing methods, including basic identification information, production context information, location information, quality assessment information, and image access information. These scattered information cannot form a comprehensive, understandable, and analyzable whole without effective integration. Therefore, by logically assembling these key data fields into a unified defect event data structure, it can be ensured that all relevant information is centrally managed and delivered.

[0034] The implementation details of step S18 can be further described as follows: the data processing module encapsulates these independent fields into a predefined data structure. This data structure can be a JSON object, an XML document, or a class instance in a programming language. For example, it can be assembled into a JSON object. This assembled JSON object is the final defect event. It contains all relevant information from the original AOI signal to the production context, location, classification, and image.

[0035] In step S2, a range query is initiated to the time series database centered on the timestamp of the defect event to obtain the defect context fingerprint. Accordingly, single defect event information is insufficient to reveal the deep reason of the defect. The occurrence of many defects is closely related to the running state of the production equipment before and after the occurrence of the defect, process parameter fluctuations, and the material batch used. In the traditional production management mode, these information is scattered in different systems, forming a data island, resulting in low defect traceability efficiency and difficulty in locating the root cause. Based on this, the present application integrates PLC real-time parameters and MES material batch information within a certain time window before and after the defect occurrence time point, which can construct a comprehensive defect context fingerprint, thereby providing rich, multi-dimensional data directly related to the defect for subsequent use.

[0036] Specifically, in an exemplary embodiment of the present application, step S2, initiating a range query to the time series database centered on the timestamp of the defect event to obtain the defect context fingerprint, comprises: S21, querying all key PLC parameters associated with the device ID in the defect event from the time series database to obtain PLC time series data; S22, querying the material batch information flowing into the device ID within 1 hour before the timestamp from the MES system; S23, data integration of the material batch information and the PLC time series data to obtain the defect context fingerprint.

[0037] The implementation details of step S2 can be further described as follows: First, step S21 is performed. The data processing module first extracts the timestamp, e.g., 2024-7-27 10:35:12.345, and the device ID, e.g., PRINTING-UNIT-002, from the defect event. Then, it initiates a query request to the time series database. Time series databases, such as InfluxDB, Prometheus, etc., are specifically designed to store data points with timestamps and can efficiently handle the writing and querying of time series data. Specifically, in an exemplary embodiment of the present application, a range query is initiated to the time series database with the timestamp of the defect event as the center, and the query range is [timestamp-15s, timestamp+2s] in the defect context fingerprint. For example, the query range is [2024-7-27 10:34:57.345, 2024-7-27 10:35:14.345]. This time window is set based on experience and understanding of the production process, that is, the direct cause of the defect often occurs in a short time before the defect occurs, and the parameter changes in a short time after the defect occurs may also provide clues. The query content includes all key PLC parameters associated with the PRINTING-UNIT-002 device ID. These key PLC parameters are pre-determined and directly affect the printing quality of the packaging bag. Specifically, in an exemplary embodiment of the present application, the key PLC parameters include the pressure, speed, tension, and drying temperature of the printing unit. Real-time data of these parameters are collected by the PLC controller and continuously written to the time series database. The query result will be all time series data points of these key PLC parameters in the specified time range, forming PLC time series data. For example, it may contain a sequence of pressure, speed, tension, and drying temperature values recorded every second.

[0038] Next, step S22 is performed. The data processing module initiates an API call to the MES system to obtain material batch information related to the defect event. The parameters of the query include the timestamp and the device ID. The query range is set to the material batch information flowing into the PRINTING-UNIT-002 device within 1 hour before the defect event timestamp. This 1 hour time window is set based on experience and aims to cover the material replacement or batch switching that may cause defects. The MES system will query its material management module according to the API request and return all material batch records flowing into the device within the time window, including material code, batch number, supplier, storage time, etc. These information constitute the material batch information.

[0039] Finally, proceed to step S23. The integration method involves structurally combining this data to form a comprehensive defect context fingerprint. For example, PLC timing data can be embedded as a time-series array, with material batch information appended as a list or dictionary. The final defect context fingerprint might be a complex JSON object containing the fluctuation trend of PLC parameters within a specific time window, as well as details of all material batches used during that time period.

[0040] In step S3, the defect context fingerprint is compared with the digital twin baseline to obtain an anomaly vector. It should be understood that even with a comprehensive defect context fingerprint, without a reference to a normal state, it is difficult to determine whether fluctuations in current production parameters constitute an anomaly. The digital twin baseline represents the expected or normal range of key PLC parameters under ideal or healthy operating conditions. By quantifying and comparing the actual PLC timing data with these baselines, it is possible to accurately identify which parameters deviated significantly before and after the defect occurred, and the degree of deviation. This quantified anomaly vector can transform complex timing data into features understandable to machine learning models, thereby providing direct anomaly signals for subsequent intelligent inference and effectively improving the accuracy and efficiency of anomaly identification.

[0041] Specifically, in one exemplary embodiment of this application, Figure 4 This is a flowchart of step S3 in the intelligent production management method for packaging bags according to an embodiment of this application. Figure 4 As shown, step S3, comparing the defect context fingerprint with the digital twin baseline to obtain an anomaly vector, includes: S31, traversing each PLC timing data in the defect context fingerprint, and performing the following operations on each PLC timing data: calculating the absolute value of the difference between the true value of each PLC key parameter in the PLC timing data and the expected value of each PLC key parameter in the digital twin baseline as the deviation value of each PLC key parameter; S32, performing statistical analysis on the time series of each PLC key parameter deviation value to obtain the comprehensive deviation score of each PLC key parameter; S33, arranging the comprehensive deviation scores of each PLC key parameter into the anomaly vector.

[0042] The implementation details of step S3 can be further described as follows: It is worth mentioning that the digital twin baseline is obtained by statistical analysis and modeling of the long-term historical data of key PLC parameters such as pressure, speed, tension, and drying temperature under normal production conditions for a specific device (e.g., PRINTING-UNIT-002). It contains the average value, standard deviation, normal fluctuation range (e.g., determined by the 3σ principle or percentile) and ideal trend pattern of each key parameter. These baseline data are established in advance by collecting, cleaning, feature engineering and statistical modeling of a large amount of defect-free production data, for example, using methods such as moving average, exponential smoothing or Kalman filtering, and stored in a queryable baseline database.

[0043] First, step S31 is performed. For each time point in the PLC time series data, the data processing module obtains the data true value of each key PLC parameter at that time point. At the same time, it obtains the data expected value of the corresponding parameter under normal conditions from the preset digital twin baseline. The data expected value can be the average value, median value of the baseline, or the ideal value predicted based on the time pattern such as periodicity, trend. Then, for each key PLC parameter, the data processing module calculates the absolute value of the difference between its data true value and data expected value. This absolute value of the difference is the PLC key parameter deviation value of the parameter at that time point. For example, if the true value of the drying temperature is 80°C at a certain time point, and the baseline expected value is 75°C, then the deviation value of the drying temperature is |80-75|=5°C. This process will iterate through all time points and all key parameters in the PLC time series data, generating a series of deviation values, i.e., the time series of each PLC key parameter deviation value.

[0044] Then, step S32 is performed. After obtaining the time series of each PLC key parameter deviation value, the data processing module will perform statistical analysis on the deviation value sequence of each parameter. The statistical analysis method can include: average deviation, calculating the average value of the deviation value time series; maximum deviation, finding the maximum value in the deviation value time series; standard deviation, measuring the volatility of the deviation value; integral deviation, integrating the deviation value within the time window, reflecting the degree of cumulative deviation; time proportion exceeding threshold, calculating the proportion of time that the deviation value exceeds a preset threshold, for example, 2 times the baseline standard deviation, in the total time window. For example, for the deviation value sequence of the drying temperature, the average deviation within a 17-second time window can be calculated, or the time proportion exceeding the 5°C threshold can be calculated. For example, in a 17-second time window, the deviation value sequence of the drying temperature is [5°C, 4°C, 6°C, 5.5°C, 7°C,...]. If the average deviation is used as the comprehensive deviation score, the average value of the sequence is calculated. If the average value is 5.8°C, the comprehensive deviation score of the drying temperature is 5.8. If the maximum deviation is used, it is 7°C. The specific statistical analysis method can be flexibly selected according to actual needs. Finally, each key PLC parameter will obtain a single comprehensive deviation score, which can summarize the abnormality degree of the parameter within the time window before and after the defect occurs.

[0045] Finally, step S33 is performed. The data processing module arranges the calculated comprehensive deviation scores of each key PLC parameter in a predetermined order to form a numerical vector. This vector is the anomaly vector. For example, if the key PLC parameters include pressure, speed, tension, and drying temperature, and their corresponding comprehensive deviation scores are [0.5, 1.2, 0.3, 0.8], then this is the final anomaly vector.

[0046] In step S4, the anomaly vector is input into a pre-trained multi-label classification machine learning model to obtain the reason probability. Accordingly, although the anomaly vector can quantify the deviation degree of the production parameters, these deviations and specific defect reasons are not a simple linear relationship, often involving complex nonlinear mapping and multi-factor coupling. For example, drying temperature anomalies can cause insufficient drying of ink, leading to trailing or sticking defects; while tension anomalies can cause material wrinkles or inaccurate overprinting. A defect can be caused by multiple reasons, or a reason can cause multiple defects. Traditional expert experience is inefficient and difficult to deal with complex situations. Based on this, in this application, a multi-label classification machine learning model is introduced to learn the complex association patterns between the anomaly vector and multiple potential defect reasons, thereby outputting the occurrence probability of each possible reason, realizing intelligent and probabilistic inference of defect root causes, and effectively improving the accuracy and efficiency of root cause analysis.

[0047] The implementation details of step S4 can be further described as follows: the multi-label classification machine learning model can adopt multiple architectures, and here a feedforward neural network (FNN) is taken as an example to illustrate, which is composed of an encoding and decoding structure. The FNN model is composed of multiple fully connected layers. The encoding structure: the input layer is designed to receive the anomaly vector, if the anomaly vector contains the comprehensive deviation score of N parameters, the input layer has N neurons. The hidden layer contains one or more fully connected layers, each layer is followed by an activation function such as ReLU, and these hidden layers are responsible for learning the complex nonlinear relationships between features in the input anomaly vector and extracting higher-level abstract features. For example, two hidden layers can be set, the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. The decoding structure: the output layer contains M neurons, where M is the number of predefined potential defect causes, each neuron corresponds to a specific defect cause, for example: drying temperature is too high, ink viscosity is abnormal, printing pressure is unstable, material batch problem, etc. The output layer uses the Sigmoid activation function, because in the multi-label classification task, a sample can belong to multiple categories at the same time, and the Sigmoid function can independently output a probability value between 0 and 1 for each category.

[0048] The training data of the model is composed of a large number of historical defect events, each event contains its corresponding anomaly vector and the true defect cause label confirmed by artificial experts. During the training process, the model continuously adjusts its internal weight and bias parameters through the backpropagation algorithm and optimizer to minimize the difference between the predicted cause probability and the true cause label. Through repeated iterative training, the model learns the complex mapping relationship between different anomaly vector patterns and various defect causes. The weights and bias parameters of the model are fixed after training is completed, and are used for subsequent inference.

[0049] In specific implementation, when the data processing module receives the anomaly vector, it sends this vector as input into the pre-loaded FNN model. The model performs the encoding process of forward propagation calculation: the anomaly vector passes through the input layer, and the data flows through each hidden layer. In each hidden layer, the input is multiplied by the weight matrix of the layer, plus the bias vector, then passes through the activation function, and outputs the anomaly encoding vector.

[0050] Decoding process: Finally, the data reaches the output layer, and each output neuron outputs a value between 0 and 1 through a sigmoid activation function. These output values are the reason probabilities for each potential defect cause. For example, the model may output a probability vector: [0.92 (dry temperature too high), 0.15 (ink viscosity abnormal), 0.08 (printing pressure unstable), 0.78 (material batch problem)]. This means that according to the current anomaly vector, the model infers that the probability of dry temperature being too high is 92%, the probability of material batch problem is 78%, and the probability of other causes is lower.

[0051] In particular, when inputting the anomaly vector into the pre-trained multi-label classification machine learning model to obtain the reason probability, the vector value based on a certain anomaly vector to obtain the reason probability based on the pre-defined label mapping relationship is only an ideal case. If you want to get the multi-value association and mapping relationship, the multi-label classification machine learning model is preferably also used in the encoder-decoder structure.

[0052] In the process of encoding the anomaly vector, considering that each vector value in the anomaly vector is a statistical analysis value of the PLC key parameter deviation value at a single time node, and due to the dimensional difference and non-uniformity of each PLC key parameter, the deviation statistical value of each time node also shows obvious time dynamic fluctuation in time series, that is, the feature value of the anomaly vector at the local relevant position after encoding There is significant statistical fluctuation, causing the decoder to decode regression instability.

[0053] Preferably, in one exemplary embodiment of the present application, the anomaly vector is input into the pre-trained multi-label classification machine learning model to obtain the reason probability, comprising:

[0054] First, the anomaly vector is input into the encoder of the pre-trained multi-label classification machine learning model to obtain an anomaly encoding vector. It should be understood that the original anomaly vector (i.e., the comprehensive deviation score arrangement of the PLC key parameters) may have a high dimension, and there is a complex non-linear relationship and redundant information between the parameters. Directly inputting the original anomaly vector into the decoder for reason inference may result in low model training efficiency and insufficient generalization ability. The role of the encoder is to map the high-dimensional original anomaly vector to an anomaly encoding vector in a low-dimensional, more abstract, and more representative encoding space or latent space. As in the above encoding structure, the encoder can be provided with two hidden layers, each containing one or more fully connected layers, and each layer is followed by an activation function such as ReLU to gradually extract the core features in the original anomaly vector and compress the redundant information to obtain the anomaly encoding vector, providing a more efficient and stable representation for subsequent optimization and decoding.

[0055] Then, the overall distance metric of the anomaly encoding vector is calculated to obtain the overall anomaly distance metric, i.e.: ;in, It is the mean of the anomaly encoding vector, and It is the number of eigenvalues ​​in the anomaly coding vector. It is the first in the anomaly coding vector Each encoded feature value, It is the overall distance metric for anomalies.

[0056] Correspondingly, even the encoded anomalous vectors may exhibit statistical fluctuations and inconsistencies in their individual encoded feature values. This stems from the dimensional differences and non-uniformity of the original PLC parameters, which are partially retained during the encoding process. Directly using these noisy encoded feature values ​​for subsequent processing may affect the regression stability of the decoder. Therefore, by calculating the overall distance metric, the overall deviation of the entire anomalous encoded vector from its average level can be quantified, providing a global anomalous strength index.

[0057] Next, based on the overall anomaly distance metric, the anomaly encoding vector is subjected to a local variation magnitude metric based on feature values ​​to obtain the anomaly local variation magnitude metric for each feature value, i.e.: ;in, The predetermined local area is determined based on experience or through cross-validation. For example, it can be set to 2, indicating that the current feature value and the next two feature values ​​are considered. It is the first in the anomaly coding vector Each encoded feature value, yes The measure of the magnitude of abnormal local changes.

[0058] It is understandable that relying solely on the overall distance metric is insufficient to capture the dynamic fluctuations and short-term state correlations of individual encoded features within anomaly encoded vectors over local time series. For example, a particular encoded feature might fluctuate dramatically within a short period, while other features remain relatively stable. This local characteristic is crucial for identifying the causes of specific defects. Therefore, a local variation amplitude metric is introduced. This metric aims to quantify the degree of significant anomaly of each encoded feature value within its local region relative to the overall anomaly distance metric. This allows for the establishment of short-term state correlations through the local variation amplitude metric. By using the short-term dependence of the vector's critical position difference to measure the significant anomalies of local temporal fluctuations relative to the overall trend, anomaly patterns can be captured more precisely, providing richer local information for subsequent establishment of global temporal fluctuation correlations.

[0059] After that, the global time fluctuation correlation is established by introducing a constraint function based on the abnormal local variation amplitude measure of each feature value to obtain an optimized abnormal encoding vector, i.e. ; wherein, is a weight hyperparameter to balance the influence of the size of the feature value itself and its local variation amplitude on optimization, which is determined in the training stage by cross-validation or grid search, for example, it can be set to 10, is a parameter function to be minimized, indicating that the goal of the entire expression is to find a set of values, so that the summation item behind it reaches a minimum value. That is, the optimization process will iteratively adjust each feature value in the abnormal encoding vector until an optimal set is found, so that the value of the entire constraint function is minimized.

[0060] Correspondingly, although the abnormal local variation amplitude measure captures short-term correlation, the fluctuations of individual encoded feature values on the time series are not completely independent, and there may be global temporal consistency or correlation between them. Directly inputting the abnormal encoding vector with local fluctuations into the decoder may cause the decoder to regress instability, affecting the model's learning of complex correlations. Therefore, by introducing a constraint function, the goal of this constraint function is to minimize a combination of the square of the encoded feature value and the square of the abnormal local variation amplitude measure, so that in the optimization process, the overall dependence of the dynamic fluctuations caused by the time node time series position deviation is unified, the time consistency of the encoded feature values is strengthened, the statistical fluctuation ambiguity under the small statistical dimension difference and statistical deviation heterogeneity is reduced, and an optimized abnormal encoding vector is obtained. While preserving key abnormal information, the unnecessary statistical fluctuations and ambiguities are eliminated, so that when input into the decoder, more accurate and stable reason probabilities can be obtained.

[0061] Finally, the optimized abnormal encoding vector is input into the decoder of the pre-trained multi-label classification machine learning model to obtain the reason probability. That is, the same as the decoding structure described above, in the decoding process, the output layer of the decoder contains multiple neurons, and uses a Sigmoid activation function to convert the optimized encoding features into probability values corresponding to each potential defect cause through the transformation of a nonlinear activation function and a weight matrix. In this way, according to the optimized abnormal encoding vector, various potential causes of defects and their occurrence probabilities are accurately inferred, thereby providing a quantitative basis for subsequent defect diagnosis and processing.

[0062] In step S5, the cause probabilities are input into a knowledge base to obtain recommended actions. It should be understood that, although the S4 step can intelligently infer the potential causes of the defect and their probabilities, these probabilities cannot directly guide the production site operators to take specific measures. Different defect causes require different intervention strategies, and these strategies are often based on production experience, process specifications, and equipment characteristics. Artificially judging and searching for corresponding solutions according to the probabilities is inefficient and prone to missing the best practices. Therefore, by combining the cause probabilities with a structured knowledge base, the present application can automatically match the most likely causes and recommend verified and targeted solutions, thereby converting the intelligent analysis results into executable action plans, effectively improving the defect response speed and problem solving efficiency.

[0063] The implementation details of step S5 can be further described as follows: The knowledge base is a structured database or knowledge graph that stores a large amount of production experience, process specifications, equipment maintenance manuals, and historical defect handling best practices. The core of the knowledge base is the mapping relationship between causes and actions, which associates each known defect cause with one or more recommended solution actions. For example, if the drying temperature is too high is the cause, the recommended actions may be to check the heating unit, adjust the temperature setting, clean the air duct, etc. At the same time, the knowledge base also sets a probability threshold for triggering the recommended actions for each cause, for example, only when the probability of a certain cause exceeds 0.7, the corresponding action is recommended. In addition, different actions can be assigned priorities or costs to facilitate sorting when there are multiple recommended actions. The data structure of the knowledge base adopts the form of relational database tables, JSON document collections, or graph databases, etc. to facilitate efficient querying and management.

[0064] In specific implementation, when the data processing module receives the cause probabilities, for example: [0.92 (drying temperature is too high), 0.15 (ink viscosity is abnormal), 0.08 (printing pressure is unstable), 0.78 (material batch problem)], it will traverse this probability vector and judge each cause. The data processing module will first filter the causes according to the preset probability threshold, for example, set the threshold to 0.7. Then, only the drying temperature is too high (0.92>0.7) and the material batch problem (0.78>0.7) will be further considered. For the causes that pass the filtering, the data processing module will use these causes as query conditions to query the knowledge base, aiming to find the recommended actions associated with these causes. The knowledge base will return all recommended actions that meet the conditions and have probabilities reaching the threshold. The data processing module can sort these actions according to the priority of the actions, the estimated effect, or the historical success rate, etc. and finally output a list of one or more recommended actions.

[0065] For example, based on the above reason probabilities, if there are the following mappings in the knowledge base: for the reason of drying temperature being too high, the recommended actions can include checking whether the drying unit heating rods are overloaded, suggesting adjusting the heating power to the set value and cleaning the drying air duct to ensure smooth ventilation and avoid heat accumulation. For the reason of material batch problem, the recommended actions can include immediately isolating the current batch of ink used (batch number: BATCH-2305-LIMEINK) and replacing it with a backup batch, and notifying the material supplier to provide a quality report of the current batch of ink. Finally, the data processing module will output a structured list of recommended actions, such as: [{"actionId":"ACT-001","description":"Check whether the drying unit heating rods are overloaded, and suggest adjusting the heating power to the set value.", "relatedReason":"Drying temperature is too high", "priority":"High"}, {"actionId":"ACT-003","description":"Immediately isolate the current batch of ink used (batch number: BATCH-2305-LIMEINK) and replace it with a backup batch.", "relatedReason":"Material batch problem", "priority":"Urgent"}, {"actionId":"ACT-002","description":"Clean the drying air duct to ensure smooth ventilation and avoid heat accumulation.", "relatedReason":"Drying temperature is too high", "priority":"Medium"}].

[0066] In step S6, the defect event, the reason probability and the recommended action are displayed on the screen in the control room. That is, all relevant information is concentrated and intuitively presented on the control room screen, enabling operators and managers to fully understand the occurrence of defects, potential causes and their likelihood at the first time, and obtain professional advice provided by the system. This visual and integrated display method greatly shortens the information transmission chain, reduces the complexity of manual judgment, ensures that production abnormalities can be discovered in time, accurately understood and efficiently handled, thereby avoiding production losses caused by information lag or decision-making errors.

[0067] The implementation details of step S6 can be further described as follows: After obtaining the recommended actions, the data processing module integrates the assembled defect event, cause probabilities, and recommended actions, and sends them to the display module in the control room through a network communication protocol such as HTTP / WebSocket. The display module in the control room is responsible for receiving these data and rendering them to the large screen or the display of the operator workstation. In specific implementation, the display module in the control room receives a complete data package containing all necessary information. The data package can be a JSON object, in which the detailed information of the defect event, the list of cause probabilities, and the list of recommended actions are nested.

[0068] After receiving this data package, the display module in the control room immediately parses its content and dynamically updates the user interface on the screen. The interface design adopts a clear and intuitive layout to present different types of information in different areas. First, in the defect event details area, the basic information of the defect event is displayed, including the defect thumbnail loaded and displayed, allowing clicking to view the large image, the defect type and severity, such as light ink smearing, severe, the time and location of occurrence, such as 2024-7-27 10:35:12, Mile Marker 28.3945, and production context information, such as Work Order No. WO-20231027-001, Product Code PROD-CHIPSBAG-LIME, Batch BATCH-2305-LIMEINK, and Equipment ID, such as PRINTING-UNIT-002. Second, in the cause probability analysis area, each potential defect cause and its corresponding probability are displayed in the form of a list or a column chart, with higher probability causes highlighted, such as displaying high drying temperature (92%) and material batch problem (78%). Finally, in the recommended action area, the system matches the recommended actions according to the cause probabilities and knowledge base, and sorts them according to priority. Each action displays its description, related cause, and priority. The operator can directly select and confirm the execution of a certain action in this area, or mark it as handled. This integrated display method allows the operators in the control room to quickly understand the causes and effects of the defect and take timely and effective intervention measures based on the professional recommendations provided by the system, thereby significantly improving the response speed and problem-solving efficiency of the production process.

[0069] In summary, the intelligent production management method for packaging bags based on the embodiments of this application is explained. When the AOI camera detects an original defect signal, the system immediately generates a defect event containing detailed production context, breaking the traditional information fragmentation. Subsequently, centered on the defect timestamp, the system queries the time series database to obtain key PLC parameters and MES material batch information, forming a comprehensive defect context fingerprint and solving the data silo problem. This fingerprint is compared with the digital twin baseline to identify anomaly vectors and reveal potential process deviations. Finally, the anomaly vectors are input into a pre-trained machine learning model, combined with a knowledge base, to intelligently infer the cause of the defect and provide accurate recommended actions, which are displayed in real time on the central control room screen. This elevates defect-cause tracing from inefficient manual experience-based judgment to efficient intelligent analysis, significantly shortening root cause location time, reducing scrap rate, and realizing intelligent and lean production management.

[0070] Figure 5 This is a block diagram of a smart production management system for packaging bags according to an embodiment of this application. Figure 5 As shown, the intelligent production management system 100 for packaging bags according to an embodiment of this application includes: a defect event generation module 110, used to generate a defect event based on the original defect signal from an AOI camera in response to the original defect signal; a defect context fingerprint acquisition module 120, used to initiate a range query to a time series database centered on the timestamp of the defect event to obtain a defect context fingerprint; an anomaly vector generation module 130, used to compare the defect context fingerprint with a digital twin baseline to obtain an anomaly vector; a cause probability generation module 140, used to input the anomaly vector into a pre-trained multi-label classification machine learning model to obtain a cause probability; a recommended action generation module 150, used to input the cause probability into a knowledge base to obtain a recommended action; and a display module 160, used to display the defect event, the cause probability, and the recommended action on a screen in the central control room.

[0071] Those skilled in the art will understand that the specific operations of each step in the above-mentioned intelligent production management system for packaging bags have been referenced above. Figures 1 to 4 The intelligent production management method for packaging bags has been described in detail, and therefore, its repeated description will be omitted.

[0072] As described above, the packaging bag intelligent production management system 100 according to the embodiments of the present application can be implemented in various wireless terminals, such as a server with a packaging bag intelligent production management algorithm, etc. In a possible implementation, the packaging bag intelligent production management system 100 according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the packaging bag intelligent production management system 100 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the packaging bag intelligent production management system 100 can also be one of the many hardware modules of the wireless terminal.

[0073] Alternatively, in another example, the packaging bag intelligent production management system 100 and the wireless terminal can also be separate devices, and the packaging bag intelligent production management system 100 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in a conventional data format.

[0074] The above has described the embodiments of the present disclosure, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A smart production management method for packaging bags, characterized in that, include: In response to the raw defect signal from the AOI camera, a defect event is generated based on the raw defect signal; Centered on the timestamp of the defect event, a range query is initiated to the time series database to obtain the defect context fingerprint; the defect context fingerprint is compared with the digital twin baseline to obtain the anomaly vector; the anomaly vector is input into a pre-trained multi-label classification machine learning model to obtain the causal probability; the causal probability is input into a knowledge base to obtain the recommended action; the defect event, the causal probability, and the recommended action are displayed on the screen in the central control room.

2. The intelligent production management method for packaging bags according to claim 1, characterized in that, In response to a raw defect signal from an AOI camera, a defect event is generated based on the raw defect signal, including: extracting the camera ID, timestamp, raw image data, defect pixel coordinates, and pulse count from the raw defect signal; generating a unique event ID; obtaining a device ID based on the camera ID through a lookup table; initiating a synchronous API call to the MES system based on the camera ID to obtain a production order signal, the production order signal including the current work order number, current product code, and current production batch; calculating the precision meter based on the pulse count; performing preliminary defect classification and severity assessment on the raw image data to obtain a preliminary defect classification result and severity score; compressing and cropping the raw image data to obtain an image-accessible URL; and assembling the unique event ID, device ID, timestamp, current work order number, current product code, current production batch, precision meter, classification result, severity score, and image-accessible URL to obtain the defect event.

3. The intelligent production management method for packaging bags according to claim 2, characterized in that, Compressing and cropping the original image data to obtain an image-accessible URL includes: compressing and cropping the original image data to obtain a thumbnail; and uploading the thumbnail to an object storage service to obtain the image-accessible URL.

4. The intelligent production management method for packaging bags according to claim 3, characterized in that, In obtaining the defect context fingerprint by initiating a range query to the time series database with the timestamp of the defect event as the center, the query range is [timestamp - 15 seconds, timestamp + 2 seconds].

5. The intelligent production management method for packaging bags according to claim 1, characterized in that, Centered on the timestamp of the defect event, a range query is initiated to the time series database to obtain the defect context fingerprint, including: querying all key PLC parameters associated with the device ID in the defect event from the time series database to obtain PLC timing data; querying the MES system for material batch information flowing into the device ID within 1 hour before the timestamp; and integrating the material batch information and the PLC timing data to obtain the defect context fingerprint.

6. The intelligent production management method for packaging bags according to claim 5, characterized in that, Key PLC parameters include the pressure, speed, tension, and drying temperature of the printing unit.

7. The intelligent production management method for packaging bags according to claim 6, characterized in that, The defect context fingerprint is compared with the digital twin baseline to obtain an anomaly vector, including: traversing each PLC timing data in the defect context fingerprint, and performing the following operations on each PLC timing data: calculating the absolute value of the difference between the true value of each PLC key parameter in the PLC timing data and the expected value of each PLC key parameter in the digital twin baseline as the deviation value of each PLC key parameter; performing statistical analysis on the time series of each PLC key parameter deviation value to obtain the comprehensive deviation score of each PLC key parameter; and arranging the comprehensive deviation scores of each PLC key parameter into the anomaly vector.

8. The intelligent production management method for packaging bags according to claim 1, characterized in that, The process of inputting the anomaly vector into a pre-trained multi-label classification machine learning model to obtain the causal probability includes: inputting the anomaly vector into the encoder of the pre-trained multi-label classification machine learning model to obtain an anomaly encoding vector; calculating the overall distance metric of the anomaly encoding vector to obtain an overall anomaly distance metric; performing a local variation amplitude metric based on feature values ​​on the anomaly encoding vector based on the overall anomaly distance metric to obtain an anomaly local variation amplitude metric for each feature value; establishing a global temporal fluctuation correlation by introducing a constraint function based on the anomaly local variation amplitude metric for each feature value to obtain an optimized anomaly encoding vector; and inputting the optimized anomaly encoding vector into the decoder of the pre-trained multi-label classification machine learning model to obtain the causal probability.

9. A smart production management system for packaging bags, characterized in that, include: A defect event generation module is used to generate a defect event based on the original defect signal from the AOI camera in response to the original defect signal. The defect context fingerprint acquisition module is used to initiate a range query to the time series database with the timestamp of the defect event as the center to obtain the defect context fingerprint; An anomaly vector generation module is used to compare the defect context fingerprint with the digital twin baseline to obtain an anomaly vector; The causation probability generation module is used to input the anomaly vector into a pre-trained multi-label classification machine learning model to obtain the causation probability. The recommended action generation module is used to input the cause probability into the knowledge base to obtain the recommended action; the display module is used to display the defect event, the cause probability, and the recommended action on the screen in the central control room.

Images