Intelligent monitoring and fault diagnosis system for bubble-free film coating process of corrugated carton

Abstract

The application belongs to the technical field of intelligent data processing, and discloses an intelligent monitoring and fault diagnosis system for a corrugated box bubble-free laminating process, which acquires full-process multi-source data and configures double data tags through a data acquisition module; a defect detection module is equipped with a lightweight deep learning model special for a laminating scene, and through a multi-scale feature fusion optimization algorithm, can identify all types of bubble defects such as micro and implicit interlayers, frame-by-frame online detection and output structured data containing defect types, positions and the like, improving the accuracy, speed and comprehensiveness of defect detection; a hierarchical early warning module establishes a three-level early warning mechanism through three core indicators, can identify the implicit risks of early parameter abnormalities, and pushes early warning data and disposal suggestions to the on-site terminal.

Description

Technical Field

[0001] This invention belongs to the field of intelligent data processing technology, specifically an intelligent monitoring and fault diagnosis system for the bubble-free lamination process of corrugated cardboard boxes. Background Technology

[0002] As a widely used packaging container, corrugated cardboard boxes undergo a lamination process that is crucial for enhancing their waterproof, moisture-proof, abrasion-resistant, and print aesthetics. Currently, quality control technologies for the lamination process of corrugated cardboard boxes mainly fall into two categories: one is machine vision-based surface defect detection technology after lamination, which uses industrial cameras to capture images of the cardboard box surface and employs traditional image processing algorithms or basic deep learning models to identify visible bubble defects; the other is process parameter monitoring technology based on fixed thresholds, which sets upper and lower limits for core lamination process parameters and triggers an alarm when these limits are exceeded. Existing technologies mostly employ end-of-line sampling or offline testing after lamination, lacking full-process online monitoring capabilities. This results in delayed detection of bubble defects, which can easily lead to the scrapping of batches of products and insufficient production cost control. Existing technologies do not deeply integrate visual defect data with multi-source data such as process operation, equipment status, environmental conditions, and material properties. Furthermore, the sampling frequency and installation location of different data sources vary, resulting in severe spatiotemporal misalignment between the data. This makes it impossible to establish a precise correspondence between defects and influencing factors. It can only identify defect phenomena but cannot locate the root cause of the failure. Existing fault diagnosis methods rely heavily on fixed thresholds set by human experience, which cannot clarify the complex relationships between multiple process parameters. They have a low recognition rate for early defects such as microbubbles and hidden bubbles, and cannot adapt to changes in working conditions such as different corrugated types, film materials, and production speeds. The rates of missed and false fault detections are high, and they cannot meet the bubble-free lamination quality requirements of high-end packaging. Existing technologies are mostly open-loop monitoring, which can only achieve defect alarms. They cannot automatically generate process parameter optimization schemes based on fault diagnosis results, nor can they achieve linkage control with the production system. Defect prevention and control rely on manual experience adjustments, resulting in low production efficiency and poor quality stability. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent monitoring and fault diagnosis system for the bubble-free lamination process of corrugated cardboard boxes, so as to solve one or more problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent monitoring and fault diagnosis system for a bubble-free lamination process for corrugated cardboard boxes, comprising the following modules: Preferably, the data acquisition module is used for online real-time synchronous acquisition and raw data encapsulation of multi-source data across the entire process of corrugated carton lamination. The acquisition scope includes all process nodes from material pretreatment before lamination, roller pressing and lamination during lamination, to cooling and shaping after lamination. The data types acquired include high-speed linear array visual image data of the carton surface during the entire lamination process, core process parameter data of the laminating machine, equipment status data of roller pressing and transmission core components, environmental condition data including workshop temperature, humidity and cleanliness, and material attribute data including corrugated board flute type and basis weight, and lamination material and thickness. At the same time, all collected data are configured with a unique encrypted dedicated workstation identifier and a hardware-level timestamp dual data tag, and the raw data is transmitted to the spatiotemporal calibration module in real time.

[0005] The dedicated workstation identifier is generated according to the 12-character rule of production line number, process node, and data acquisition device number. It is encrypted using MD5 to ensure that the identifier cannot be tampered with. The hardware-level timestamp is generated based on the pulse signal of the main shaft encoder of the laminating production line, with a time accuracy of milliseconds. It is synchronized with the hardware clock of the data acquisition device. Each piece of acquired data is bound to a unique encrypted workstation identifier and a millisecond-level timestamp, forming a one-to-one dual data tag. The tag and the original data are encapsulated in a unified data format for transmission.

[0006] The key process parameters for equipment operation are collected, including roller temperature, roller pressure, laminating speed, adhesive coating amount, and unwinding / rewinding tension. The key component status data are monitored, including roller speed, motor current / voltage, bearing temperature, and drive belt tension. The workshop environment data are continuously collected, including temperature, relative humidity, air cleanliness, and workshop air pressure. The material property data are accurately collected, including corrugated cardboard flute type, basis weight, moisture content, laminating material, thickness, and light transmittance. All collected indicators are set within the normal process range.

[0007] For the two core processes of roll forming and cooling and setting, parameter control details related to bubble defects were established. The roll forming process adopts dual-roll synchronous roll forming control to ensure that the temperature deviation between the upper and lower rolls is ≤5℃ and the pressure deviation is ≤1MPa. The roll forming speed is matched with the unwinding / rewinding tension in real time. When the tension fluctuation exceeds ±2N, the parameter linkage adjustment is immediately triggered to avoid the formation of bubbles due to film stretching / relaxation. The cooling and setting process adopts segmented temperature control. The cooling inlet temperature is set to 40-60℃, and the deviation between the cooling outlet temperature and the workshop ambient temperature is ≤10℃. The cooling wind speed is controlled at 2-5m / s to prevent the formation of bubbles due to thermal expansion and contraction after lamination. The real-time values ​​and deviation values ​​of each control parameter are included in the equipment operation parameter acquisition system and are simultaneously bound to dual data tags. As the core influencing factor of bubble defects, the adhesive coating process requires precise collection of four key parameters: coating amount, coating uniformity, adhesive solid content, and coating speed. The coating amount is matched to the paperboard weight at 10-20 g / m², the coating uniformity deviation is ≤5%, and the adhesive solid content is controlled between 30% and 50%. The coating speed is synchronized with the lamination speed at a 1:1 ratio to avoid uneven adhesive coating that can cause bubbles. At the same time, the temperature and viscosity data of the adhesive are collected. The adhesive temperature is maintained at 25-35℃, and stirring / temperature control adjustment is triggered when the viscosity fluctuation exceeds ±50 mPa・s. All coating parameters are collected in conjunction with the rolling parameters to form a full-link parameter correlation between coating and rolling.

[0008] Targeted noise preprocessing was performed on the collected raw data. For visual image data, median filtering was used to eliminate salt-and-pepper noise generated by industrial camera shooting, while retaining the edge features of bubble defects. For equipment operating parameters and environmental condition data, moving average filtering was used, with 5 sampling points as a sliding window to remove sudden pulse noise. For core component status data, amplitude limiting filtering was used to identify and remove outliers that exceeded the rated parameter range of the equipment. The preprocessed data maintained the original time sequence and label association, with no feature loss or data misalignment issues.

[0009] Preferably, the spatiotemporal calibration module receives the raw multi-source data transmitted by the data acquisition module, performs spatiotemporal synchronization, format standardization and matching of multi-source heterogeneous data, and the spatiotemporal calibration module uses the main shaft encoder of the coating production line to synchronize the pulse signal, builds a hardware-level timestamp alignment algorithm, and completes the unified timing calibration of data acquired at different sampling frequencies. By combining the spatial coordinate system of the coating production line with the workstation identification data tags, a one-to-one correspondence is established between the location of visual defects, the corresponding process workstation, and the corresponding acquisition parameters. The spatiotemporal calibration module is equipped with a spatiotemporal misalignment correction algorithm and a data fusion verification mechanism to perform consistency verification on the calibrated data, remove invalid and misaligned data fragments, and output a standardized dataset that is spatiotemporally synchronized. The dataset is transmitted to the defect detection module and the root cause diagnosis module in real time.

[0010] The hardware-level timestamp alignment algorithm uses 1000 pulses / revolution of the production line spindle encoder as a benchmark. It maps data from different sampling frequencies of visual image acquisition devices (20 frames / second), process parameter acquisition sensors (1 time / second), and environmental monitoring devices (1 time / 5 seconds) onto the time axis of the benchmark pulse, achieving time sequence unification of different data sources. The spatiotemporal misalignment correction algorithm compares the coordinates of the defect location in the visual image with the physical coordinates of the corresponding process station through spatial coordinate matching verification. If the deviation exceeds 5mm, it is judged as spatial misalignment and coordinate correction is performed. The data fusion verification mechanism adopts dual-dimensional verification of data integrity and consistency. Data with missing labels or discontinuous timestamps is judged as invalid data. Parameter values ​​that exceed the normal process range and have no corresponding image data are judged as inconsistent data. Invalid and inconsistent data are directly discarded.

[0011] Preferably, the defect detection module receives spatiotemporally synchronized standardized visual image data output by the spatiotemporal calibration module and performs online, full-type data recognition of lamination bubble defects throughout the entire process. The defect detection module has a built-in lightweight deep learning detection model specifically for lamination scenarios. For all types of defects in the corrugated cardboard box lamination process, such as micro bubbles, hidden interlayer bubbles, edge bubbles, and pinhole bubbles, a refined defect feature extraction algorithm is built through multi-scale feature fusion and channel attention mechanism optimization. The defect detection module simultaneously simplifies the redundant structure of the general target detection model, adapts to the real-time inference requirements of the coating production line, completes frame-by-frame online detection of visual images of the coating process, and synchronously outputs structured data results including defect type, size, quantity, spatial location, and severity. The data results are transmitted to the root cause diagnosis module and the graded early warning module in real time.

[0012] The lightweight deep learning detection model is based on the general YOLO model. The core layers include a lamination image input layer, a multi-scale feature extraction layer, and a channel attention mechanism layer that integrates 3×3, 5×5, and 7×7 convolutional kernels to adapt to the texture features of the corrugated cardboard lamination surface. The channel attention mechanism layer assigns weight coefficients of 0.6-0.9 to key feature channels such as bubble edges and interlayer dark patterns. The model input is high-speed linear array visual image data of the 512×512 resolution corrugated cardboard lamination surface after spatiotemporal calibration. The image acquisition frame rate is matched with the production line speed at 5-20 frames / second. The model output is structured defect data bound to the spatial coordinates of the lamination production line. Each output data field is associated with the workstation identifier and timestamp label. The severity of bubble defects is judged on four levels: slight, mild, moderate, and severe. Slight defects are pinhole bubbles (diameter <0.5mm) with fewer than 5 bubbles per frame and no hidden layered bubbles. Mild defects are tiny bubbles (diameter 0.5-2mm) with 5-20 bubbles per frame or edge bubbles with a length <5cm and hidden layered bubble area <1cm². 2 Moderate defects are characterized by bubbles with a diameter of 2-5mm, numbering 10-30 per frame, or edge bubbles with a length of 5-15cm, and hidden interlayer bubbles with an area of ​​1-5cm². 2 Severe defects are defined as bubbles with a diameter > 5 mm or more in a single frame than 30, edge bubble length > 15 cm, or hidden interlayer bubble area > 5 cm². 2 The judgment result is directly mapped to a severity coefficient of 0-1, which serves as the core input data for the graded early warning model.

[0013] Preferably, the root cause diagnosis module synchronously receives standardized full-dimensional process, equipment, environment, and material parameter data output by the spatiotemporal calibration module and defect structured data output by the defect detection module, and performs intelligent root cause localization of bubble defects. The root cause diagnosis module has a built-in causal structure learning model, which completes deep learning training based on historical production data to construct a causal relationship network between bubble defects and various influencing parameters. A multi-parameter correlation decomposition analysis method was established to clarify the complex relationships between multiple process parameters; the core root cause parameters that induce bubble defects were located, and their influence weights and risk levels were quantitatively determined, outputting structured diagnostic data. The data was transmitted in real time to the graded early warning module, the process optimization module, and the data closed-loop iteration module.

[0014] The causal structure learning model takes multi-source data of the lamination process as input. The input dimensions include four major categories of feature parameters: equipment operating parameters, environmental condition parameters, material property parameters, and defect structured data. Equipment operating parameters include rolling temperature, pressure, and speed; environmental condition parameters include temperature, humidity, and cleanliness; and material property parameters include paperboard basis weight and film thickness. The model constructs a causal relationship network using the PC algorithm. The network nodes are various process parameters and bubble defect types, and the edge weights are the influence coefficients of the parameters on the defects (values ​​range from 0 to 1). During model training, data from over 1000 batches of coating processes in historical production are used as samples. The network node relationships are optimized through Bayesian estimation. The model output is structured diagnostic data consisting of core root cause parameter names, influence weight values, and risk levels (low, medium, and high). Each output parameter is bound to its corresponding process station and timestamp label.

[0015] The multi-parameter correlation decomposition analysis method is based on the process logic. It decomposes the 28 influencing parameters of the entire coating process into three categories: material properties, equipment operation, and environmental conditions. Then, it decomposes them into two categories: pre-coating, mid-coating, and post-coating process nodes. For the decomposed parameter subsets, the correlation between single parameter changes and the occurrence of bubble defects is analyzed one by one. Then, the comprehensive impact of the coupled changes of multiple parameters in the same process and across processes on defects is analyzed. Redundant parameters that are not related to bubble defects and have a correlation coefficient of less than 0.1 are eliminated. Finally, the core root cause parameter subset is identified.

[0016] Preferably, the graded early warning module receives defect detection data transmitted by the defect detection module and root cause quantitative diagnostic data output by the root cause diagnosis module, and performs bubble defect risk graded early warning and pre-control operations. The graded early warning module has a built-in data-driven graded early warning model, which divides the early warning mechanism into three levels based on three data indicators: defect severity, root cause risk weight, and defect propagation probability. The first level is a hidden risk warning, which corresponds to early data abnormality conditions where parameters deviate from the normal range and there are no visible defects. The second level is a local abnormality warning, which corresponds to minor defects and local data abnormality conditions. The third level is a batch failure warning, which corresponds to severe defects and global data abnormality conditions. Each level of early warning is matched with a differentiated data control strategy. The graded early warning module completes the early warning grade based on data analysis throughout the process. The early warning signal and graded data are transmitted to the process optimization module in real time, and the early warning data and handling suggestions are pushed to the on-site control terminal at the same time.

[0017] The data-driven hierarchical early warning model takes three core feature data as input: the defect severity coefficient output by the defect detection module, the root cause parameter risk weight output by the root cause diagnosis module, and the defect propagation probability (valued at 0-1) calculated based on Markov chains. The defect severity coefficient, with a value of 0-1, is calculated by comprehensively considering defect size, quantity, and location. The model constructs an early warning classification judgment model using a logistic regression algorithm, setting judgment thresholds: Level 1 latent risk early warning is defined as a defect severity coefficient = 0, a root cause risk weight ≥ 0.3, and a propagation probability < 0.5; Level 2 local anomaly early warning is defined as a defect severity coefficient 0.1-0.5, a root cause risk weight ≥ 0.5, and a propagation probability 0.5-0.8; and Level 3 batch failure early warning is defined as a defect severity coefficient > 0.5, a root cause risk weight ≥ 0.7, and a propagation probability > 0.8. The model output includes the early warning level, the corresponding control strategy code, and the handling suggestion code, with the output data format adapted to the display interface data of the production line workstation terminal.

[0018] Early warning data and handling suggestions are pushed to the on-site control terminals via the Industrial Internet of Things (IIoT) protocol. The terminals are divided into the main control terminal of the production line and the on-site terminals of each process station. The first-level hidden risk warning is only pushed to the main control terminal, which displays the abnormal parameter name, corresponding station, and risk weight in a yellow pop-up window, along with parameter fine-tuning suggestions. The second-level local anomaly warning is pushed to both the main control terminal and the corresponding process station terminal, which displays the defect information, root cause parameters, and local control strategies in an orange pop-up window, and supports one-click access to real-time image data at the station terminal. The third-level batch fault warning is pushed to all terminals on the entire production line, which displays the fault information, defect propagation probability, and emergency stop / speed reduction suggestions in a red pop-up window. All pop-ups include confirmation and receipt buttons. The terminal operation records are transmitted back to the hierarchical warning module in real time, forming a closed loop of warning push-handling-receipt.

[0019] Preferably, the process optimization module receives the warning signal and graded data transmitted by the graded warning module and the fault root cause diagnosis data output by the root cause diagnosis module, and performs adaptive closed-loop optimization and intelligent decision-making of the coating process parameters. The process optimization module has a built-in process parameter adaptive optimization model based on deep reinforcement learning, with bubble-free coating as the core optimization goal and production efficiency and material loss control as data constraints, and automatically generates process parameter adjustment schemes. The process optimization module uses data processing logic to deduce and generate optimization schemes, which are then simultaneously sent to the production line control system.

[0020] The deep reinforcement learning model uses the real-time operating conditions of the lamination process as its state space, which includes real-time values ​​of adjustable parameters such as roller temperature (80-180℃), roller pressure (5-20MPa), lamination speed (10-50m / min), workshop temperature (20-30℃), and workshop humidity (30%-60%). The action space uses the adjustment step size for each parameter: 5℃ for temperature, 1MPa for pressure, and 2m / min for speed. The core reward function is bubble-free lamination, with a positive... Positive and negative rewards are assigned based on the severity of defects. A reward of +10 is given when there are no bubbles, and -5 to -20 is given when bubbles appear. Production efficiency ≥30 m / min and material loss ≤1% are used as constraints; if these constraints are not met, the reward is reduced by 5. The model inputs include tiered early warning data, root cause diagnosis data, and real-time operating parameters. The output is a structured optimization scheme containing the target values, adjustment directions, and adjustment steps for each adjustable process parameter. The output data format is consistent with the parameter input format of the production line PLC control system.

[0021] The built-in data processing logic verification is executed in three steps: The first step is parameter rationality verification, which checks whether the target values ​​of process parameters in the optimization plan are within the safe operating range of the equipment and the standard range of the coating process. If they exceed the range, they are automatically corrected to the critical value of the range. The second step is working condition adaptability verification, which matches the optimization plan with the actual working conditions of the current coating production line, such as flute type, film material, and production speed, to ensure that the parameter adjustment is adapted to the current production conditions. The third step is effect pre-verification, which extrapolates the expected elimination rate of bubble defects after the optimization plan is implemented based on historical process data. If the expected elimination rate is less than 80%, the optimization plan is re-generated. Only after the optimization plan passes the verification can it be sent to the production line control system.

[0022] Dedicated process adaptation logic is built for different corrugated flute types and laminating materials. For A / C type coarse corrugated cardboard, higher rolling pressure and slower laminating speed are adapted to avoid the flute tips puncturing the film and causing air bubbles. For B / E type fine corrugated cardboard, lower rolling pressure and faster laminating speed are adapted to prevent the cardboard from being crushed and causing poor lamination. For BOPP film, a rolling temperature of 80-120℃ is adapted, and for PET film, a rolling temperature of 130-180℃ is adapted. The model has a built-in flute type and film material process parameter adaptation table. After collecting material attribute data in real time, it automatically matches the benchmark parameters in the table and then makes fine adjustments based on root cause diagnosis data to achieve accurate adaptation for different working conditions. The system's core AI models are all deployed on the production line edge computing server, which has 8 or more CPU cores, an industrial-grade inference GPU, and ≥32GB of memory. It is directly connected to the field acquisition equipment and control terminals via industrial Ethernet. The defect detection model is deployed separately on the vision inspection substation, and the model inference process is completed locally at the edge, with inference latency controlled within 200ms to meet the real-time requirements of the production line. It is directly connected to high-speed linear industrial cameras to realize localized processing of image acquisition and defect detection. The remaining models are integrated and deployed on the main edge computing server. The models interact with each other through standardized data interfaces using the TCP / IP protocol and JSON data transmission format to ensure efficient and compatible data interaction between models. All models support local offline operation to avoid network failures affecting normal production line monitoring.

[0023] The process optimization module and the production line PLC control system use the standard Modbus-RTU protocol for instruction interaction. The optimization scheme is converted into digital instructions that the PLC can recognize. The instructions include four core fields: parameter type code, target value, adjustment step size, and execution time. After the instruction is issued, the PLC control system provides real-time feedback on the instruction reception status. If the reception is successful, an acknowledgment code is returned; if the reception fails, an error code is returned with the reason. After the PLC performs parameter adjustment, it feeds back the actual executed value of the parameter to the process optimization module every second, continuously feeding back until the parameter is stable within ±2% of the target value, thus completing the execution feedback. If the parameter exceeds the equipment's safe range during execution, the PLC immediately stops adjusting and sends an emergency feedback signal to the process optimization module. Upon receiving the signal, the process optimization module immediately re-derives and generates the optimization scheme.

[0024] Preferably, the data closed-loop iteration module receives full-process operation data transmitted by the data acquisition module, spatiotemporal calibration module, defect detection module, root cause diagnosis module, graded early warning module, and process optimization module, including raw acquisition data, standardized data after calibration, defect detection data, root cause diagnosis data, early warning data, process optimization data, and control effect feedback data. The data closed-loop iteration module builds a standardized dataset specifically for the coating process, completes the entire process of data cleaning, labeling, classification, storage and compliance management, and has a built-in incremental learning mechanism. Based on newly added production data and control feedback data, it performs adaptive incremental iterative training for the system's defect detection, root cause diagnosis and process optimization models. At the same time, it is configured with a full-process data traceability algorithm to build a dedicated production data ledger for each batch of products.

[0025] The end-to-end data traceability algorithm uses the product production batch number as the core traceability identifier. It collects data from all processes of a single batch of products, from before to after lamination, including calibration data, defect detection data, root cause diagnosis data, early warning data, process optimization data, and control effect feedback data. These data are then bound to the batch number through dual data tags, forming a traceability link of batch number-tag-end-data. When querying traceability, entering the batch number will retrieve all related data for that batch of products. It supports forward traceability by process and reverse traceability by defect / early warning results, with traceability accuracy down to specific workstations and specific time nodes.

[0026] The cleaning of the entire process operation data follows standardized execution rules. First, missing value handling is performed. When the missing field of a single data entry is ≤20%, it is filled with the average value of the same batch and workstation. If the missing field is >20%, it is directly removed. Second, outlier handling is performed. The quartile method is used to determine outlier parameters. Values ​​exceeding 1.5 times the upper and lower quartile interval are marked as outliers and reviewed. Data without actual process fluctuations are directly removed. Finally, duplicate data handling is performed. Duplicate data is determined based on dual data labels. The earliest timestamp is retained and the rest of the duplicate data is deleted. The cleaned data is labeled with a cleaning mark and included in a dedicated standardized dataset. Feedback data on the control effect is collected starting 5 seconds after the process optimization plan is issued, with 30 seconds of continuous collection of full-dimensional data on the coating process as feedback samples. The control effect is judged from three dimensions: defect elimination, parameter stability, and production efficiency. For defect elimination, the reduction percentage of the severity coefficient of bubble defects after optimization is used as the basis: a reduction ≥80% is considered excellent, 50%-80% is considered good, and <50% is considered poor. For parameter stability, the fluctuation range of the adjustable process parameters is used as the basis: fluctuation ≤5% is considered stable, otherwise it is considered unstable. For production efficiency, whether the coating speed and material loss meet the constraints is used as the basis; if they do, it is considered satisfactory. If all three dimensions meet the standards and defect elimination is excellent or good, the control is considered effective; otherwise, it is considered ineffective. The judgment results are linked to the feedback samples and included in the incremental learning dataset.

[0027] The entire process of multi-source data is classified and stored in three categories: raw data, processed data, and result data. Raw data includes unprocessed data such as visual images, parameters, and statuses with dual labels output by the acquisition module, stored in its original format on a large-capacity hard drive, with a retention period of 2 years. Processed data includes standardized datasets after spatiotemporal calibration and incremental learning samples after data cleaning, stored in a structured database format (MySQL), with secondary indexes built according to production batches and process nodes, and a retention period of 3 years. Result data includes all output results of defect detection, root cause diagnosis, graded early warning, process optimization, and control effect judgment data, stored in a lightweight database format (Redis), and simultaneously synchronized to the cloud for backup and permanent storage. All data storage is encrypted in accordance with industrial data security standards, and only authorized personnel can access it through production batch number / workstation identifier, ensuring data security and confidentiality.

[0028] The beneficial effects of this invention are as follows: 1. The system of this invention acquires multi-source data from the entire process through the data acquisition module and configures dual data tags; the defect detection module is equipped with a lightweight deep learning model specifically for the coating scenario. Through multi-scale feature fusion optimization algorithm, it can identify all types of bubble defects such as small and hidden interlayers. The simplified model structure adapts to the real-time reasoning needs of the production line. It detects and outputs structured data containing defect type, location, etc. frame by frame online, replacing the subjective judgment and lag of manual inspection, and improving the accuracy, speed and comprehensiveness of defect detection.

[0029] 2. The root cause diagnosis module of this invention constructs a causal network of defects and parameters based on a causal structure learning model, quantitatively analyzes the influence weight and risk level of core root cause parameters, and clarifies the complex relationship of multiple parameters; the graded early warning module establishes a three-level early warning mechanism through three core indicators, which can identify the hidden risks of early parameter anomalies and push early warning data and handling suggestions to the field terminal.

[0030] 3. The process optimization module of this invention is based on a deep reinforcement learning model, which automatically generates optimization schemes and sends them to the production line control system; the data closed-loop iteration module gathers the full-process operation data to build a dedicated dataset, and continuously iterates and optimizes the core model through an incremental learning mechanism to improve the model's adaptability to multiple working conditions. At the same time, it configures a data traceability algorithm to establish a ledger for each batch of products, reduce material loss, and continuously improve the stability and process level of the coating production. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall operation of the system of the present invention; Figure 2 This is a flowchart of the root cause diagnosis of bubble defects in this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] like Figures 1 to 2 As shown, this embodiment of the invention provides an intelligent monitoring and fault diagnosis system for the bubble-free lamination process of corrugated cardboard boxes, including the following modules: The data acquisition module is used to collect and pre-process multi-dimensional, multi-source data of the entire process of corrugated cardboard box lamination in real time, including key process nodes of the entire lamination process. At the same time, it configures a unique encrypted dedicated workstation identifier and hardware-level timestamp dual data tag for all collected data, and transmits the raw data to the spatiotemporal calibration module in real time. The spatiotemporal calibration module receives raw multi-source data transmitted by the data acquisition module, completes spatiotemporal synchronization calibration and format standardization processing of various heterogeneous data, corrects spatiotemporal misalignment deviations of data, establishes a unified correspondence between various types of data, and outputs a standardized dataset with spatiotemporal synchronization after consistency verification. The defect detection module receives standardized visual image data output by the spatiotemporal calibration module, is equipped with a special intelligent detection model for the coating scenario, extracts the core features of defects and generates structured detection data, and outputs the defect judgment results in real time. The root cause diagnosis module simultaneously receives standardized full-dimensional operating condition parameter data and defect detection data. Based on the built-in intelligent diagnostic model, it completes the root cause localization of bubble defects, performs quantitative analysis of root cause parameters, influence weights and risk levels, and outputs structured diagnostic data. The graded early warning module receives defect detection and root cause diagnosis data, completes defect risk classification based on core data indicators, generates corresponding early warning signals and control instructions, and outputs early warning information simultaneously. The process optimization module receives graded early warning data and root cause diagnosis data, generates process parameter adjustment schemes adapted to real-time operating conditions based on the built-in intelligent optimization model, and sends optimization instructions to the production line control system. The data closed-loop iteration module receives process data from the aforementioned modules, builds a standardized dataset specifically for the coating process, and completes adaptive iterative optimization of the system's core model.

[0034] The data acquisition module is used for online real-time synchronous acquisition and raw data encapsulation of multi-source data across the entire process of corrugated carton lamination. The acquisition scope includes all process nodes from material pretreatment before lamination, roll forming during lamination, to cooling and shaping after lamination. The data types acquired include high-speed linear array visual image data of the carton surface during the entire lamination process, core process parameter data of the laminating machine, equipment status data of roll forming and transmission core components, environmental condition data including workshop temperature, humidity, and cleanliness, and material attribute data including corrugated board flute type and basis weight, and lamination material and thickness. Simultaneously, all collected data are configured with a unique encrypted dedicated workstation identifier and a hardware-level timestamp dual data tag, completing the formatted encapsulation, noise preprocessing and time-series sorting of the raw data, and transmitting the raw data to the spatiotemporal calibration module in real time.

[0035] The dual data tags and the original data are encapsulated in JSON data format, with the tag fields set at the top. The workstation identification encryption field is named Encrypt_StationID, and the hardware-level timestamp field is named Hardware_Timestamp. After the encapsulated data is verified to have no missing tags or incorrect timestamps, it is pushed to the spatiotemporal calibration module in real time via industrial Ethernet at a transmission rate of 100Mbps to ensure that the tags and the original data are not separated during data transmission.

[0036] The acquisition devices and acquisition frequencies for different types of data are matched as follows: high-speed linear scan industrial cameras acquire visual image data at a frequency of 20 frames / second; pressure and temperature sensors acquire process parameters of the equipment at a frequency of 1 time / second; current and temperature detectors acquire status data of core components at a frequency of 1 time / 2 seconds; temperature, humidity, and cleanliness sensors acquire environmental operating condition data at a frequency of 1 time / 5 seconds; and material attribute data is acquired by barcode scanning equipment, which is acquired once before lamination and synchronized to the full-process data, so as to achieve precise matching between acquisition frequency and process rhythm.

[0037] The spatiotemporal calibration module receives the normalized multi-source data transmitted by the data acquisition module, performs spatiotemporal synchronization, format standardization and matching of multi-source heterogeneous data, and builds a hardware-level timestamp alignment algorithm through the main shaft encoder of the coating production line to complete the unified timing calibration of the acquired data at different sampling frequencies. By combining the spatial coordinate system of the coating production line with the workstation identification data tags, a one-to-one correspondence is established between the location of visual defects, the corresponding process workstation, and the corresponding acquisition parameters, eliminating spatiotemporal misalignment of data and realizing the fusion of multiple types of heterogeneous data. The spatiotemporal calibration module is equipped with a spatiotemporal misalignment correction algorithm and a data fusion verification mechanism to perform consistency verification on the calibrated data, eliminate invalid and misaligned data fragments, and output a standardized dataset that is spatiotemporally synchronized. The dataset is transmitted to the defect detection module and the root cause diagnosis module in real time, ensuring the accuracy of fault root cause location from the data layer.

[0038] The specific steps for performing spatiotemporal calibration are as follows: 1. Receive the encapsulated raw multi-source data and parse the workstation identifier and timestamp in the dual data tags; 2. A hardware-level timestamp alignment algorithm is used to unify the time series of all data sources, generating a dataset with a unified timeline; 3. Combining the production line spatial coordinate system, a spatiotemporal misalignment correction algorithm is used to complete the spatial matching and coordinate correction of defect locations with workstations and parameters; 4. Activate the data fusion verification mechanism to verify the integrity and consistency of data batch by batch, and remove invalid or misaligned data fragments; 5. The calibrated dataset is categorized and formatted into visual image data and operating condition parameter data, and a standardized dataset with spatiotemporal synchronization is output. The two types of datasets can be linked for querying through the same label field.

[0039] The defect detection module receives spatiotemporally synchronized standardized visual image data output by the spatiotemporal calibration module and performs online, full-type data recognition of lamination bubble defects throughout the entire process. The defect detection module has a built-in lightweight deep learning detection model specifically for lamination scenarios. For all types of defects in the corrugated cardboard box lamination process, such as micro bubbles, hidden interlayer bubbles, edge bubbles, and pinhole bubbles, it builds a refined defect feature extraction algorithm through multi-scale feature fusion and channel attention mechanism optimization, thereby improving the feature recognition accuracy and data processing speed of early hidden defects. The defect detection module simultaneously simplifies the redundant structure of the general target detection model, adapts to the real-time inference requirements of the coating production line, completes frame-by-frame online detection of visual images of the coating process, and synchronously outputs structured data results including defect type, size, quantity, spatial location, and severity. The data results are transmitted in real time to the root cause diagnosis module and the graded early warning module to realize real-time analysis and synchronous control of defect data.

[0040] The training samples for the target detection model consist of images of all types of bubble defects (A / B / C / E type) and different film materials (BOPP, PET) in the corrugated cardboard lamination process, with a sample size of no less than 100,000 images. During training, the batch size was set to 32, the learning rate to 0.001, and the number of iterations to 200. The model parameters were optimized using the MSE loss function. Oversampling was performed on difficult-to-identify samples such as microbubbles and hidden interlayer bubbles, with a sampling ratio of 1:4, to ensure the recognition accuracy of the model in the actual working conditions of the lamination production line.

[0041] The root cause diagnosis module synchronously receives standardized full-dimensional process, equipment, environment, and material parameter data output by the spatiotemporal calibration module and defect structured data output by the defect detection module, and performs intelligent root cause localization of bubble defects. The root cause diagnosis module has a built-in causal structure learning model, which completes deep learning training based on historical production data to construct a causal relationship network between bubble defects and various influencing parameters. A multi-parameter correlation decomposition analysis method was established to clarify the complex correlation between multiple process parameters; the core root cause parameters that induce bubble defects were located, and their influence weight and risk level were quantitatively determined. Structured diagnostic data was output, and the data was transmitted in real time to the graded early warning module, the process optimization module, and the data closed-loop iteration module to complete the whole process data analysis from defect phenomenon identification to root cause quantitative location.

[0042] The training steps for the causal structure learning model are as follows: Historical lamination process data was cleaned, and batches with missing values ​​or outliers exceeding 30% were removed. The input feature parameters are standardized to map the parameter values ​​to the interval [0, 1]. A causal network of defects and parameters is constructed using the PC algorithm to determine the core root cause parameter set; Using the defect occurrence rate as a label, the gradient boosting tree algorithm is used to quantitatively calculate the influence weight of each root cause parameter; Risk levels are determined based on weight values: ≥0.7 indicates high risk, 0.3-0.7 indicates medium risk, and <0.3 indicates low risk. This completes model training and parameter calibration.

[0043] The specific implementation steps of multi-parameter correlation decomposition analysis are as follows: It receives standardized full-dimensional parameter data and defect structured data, and performs two-level decomposition according to parameter category and process node to form several parameter subsets; Perform single-parameter correlation analysis on each parameter subset and statistically analyze the defect incidence rate when the parameters deviate from the normal range; Coupled correlation analysis was performed on parameters with high occurrence rates in the same process, and the defect occurrence rate and defect severity were statistically analyzed when multiple parameters deviated simultaneously. Set the correlation coefficient threshold to 0.1, eliminate redundant parameters, and lock in the core root cause parameters; The core root cause parameters are input into the causal structure learning model to complete the subsequent quantitative determination of influence weights and risk levels.

[0044] The graded early warning module receives defect detection data transmitted by the defect detection module and root cause quantitative diagnostic data output by the root cause diagnosis module, and performs bubble defect risk graded early warning and pre-control operations. The graded early warning module has a built-in data-driven graded early warning model, which divides the early warning mechanism into three levels based on three data indicators: defect severity, root cause risk weight, and defect propagation probability. The first level is a hidden risk warning, which corresponds to early data abnormality conditions where parameters deviate from the normal range and there are no visible defects. The second level is a local abnormality warning, which corresponds to minor defects and local data abnormality conditions. The third level is a batch failure warning, which corresponds to serious defects and global data abnormality conditions. Each level of early warning is matched with a differentiated data control strategy. The graded early warning module completes the early warning classification based on data analysis throughout the entire process. The early warning signal and graded data are transmitted to the process optimization module in real time, and the early warning data and disposal suggestions are pushed to the on-site control terminal at the same time, realizing the whole process quality control of proactive data prevention and control.

[0045] The data-driven hierarchical early warning model was trained using historical early warning data and fault handling data of the coating process as samples. The samples included more than 500 first-level early warnings, more than 300 second-level early warnings, and more than 200 third-level early warnings. During training, a regularization coefficient of 0.01 was set to avoid overfitting. The model effect was verified by two indicators: accuracy and recall. In practical applications, the model can receive input data in real time and complete the early warning classification within 100ms. The output results are directly synchronized to the database of the production line control terminal.

[0046] The process optimization module receives warning signals and graded data transmitted by the graded warning module and fault root cause diagnosis data output by the root cause diagnosis module. It performs adaptive closed-loop optimization and intelligent data decision-making of coating process parameters. The process optimization module has a built-in process parameter adaptive optimization model based on deep reinforcement learning. It takes bubble-free coating as the core optimization goal and production efficiency and material loss control as data constraints. Combined with real-time operating data, it automatically generates process parameter adjustment schemes. The process optimization module uses data analysis and deduction logic based on the built-in intelligent optimization model to complete the deduction and generation of optimization schemes. The optimization schemes are simultaneously sent to the production line control system, establishing a closed-loop data management system for defect detection, root cause diagnosis, parameter optimization, and process control. No manual intervention is required throughout the process, eliminating bubble defects and blocking defect propagation from the data level.

[0047] The intelligent optimization model is trained using the DDPG algorithm. The training agent is a decision-making unit for controlling the coating process parameters. The training steps are as follows: 1. Initialize the experience pool and store 10,000+ data entries of parameter adjustments and results under different working conditions of the coating process; 2. Set the exploration rate ε=0.9, and gradually decrease it to 0.1 with training iterations to balance model exploration and utilization; 3. The intelligent agent outputs parameter adjustment actions and interacts with the coating process simulation environment to obtain reward values ​​and new working conditions; 4. Store the interaction data in the experience pool, and randomly sample batch data to update the parameters of the Actor network and Critic network; 5. When the average reward value of the model stabilizes above 8 for 100 consecutive iterations, the training is complete and the model is deployed to the actual production line. In practical applications, the model can update the parameter optimization scheme every 5 seconds based on real-time changes in operating conditions.

[0048] The specific process for verifying and issuing the optimization plan is as follows: 1. The deep reinforcement learning model generates an initial process parameter optimization scheme, which is then passed to the built-in data processing logic verification unit; 2. Perform parameter rationality verification, and check whether the target values ​​of parameters such as roller pressing temperature, pressure, and speed are within the equipment and process standard range; 3. Perform operating condition adaptability verification, match the current material properties, equipment status, and environmental conditions, and adjust the step size of any incompatible parameters; 4. Perform pre-validation of the execution effect, and extrapolate the expected defect elimination rate based on historical data. If the rate is less than 80%, return to the model and regenerate the solution. 5. After all three verifications are passed, the optimized solution is converted into an instruction format that the production line control system can recognize, and sent to the PLC control system of the laminating machine in real time via the industrial bus to realize intelligent control of process parameters.

[0049] The data closed-loop iteration module receives full-process operation data transmitted from the data acquisition module, spatiotemporal calibration module, defect detection module, root cause diagnosis module, graded early warning module, and process optimization module, including raw acquisition data, standardized data after calibration, defect detection data, root cause diagnosis data, early warning data, process optimization data, and control effect feedback data. The data closed-loop iteration module builds a standardized dataset for intelligent management and control of the coating process, completing data cleaning, labeling, classification, storage, and compliance management throughout the entire process. It has a built-in incremental learning mechanism that performs adaptive incremental iterative training on defect detection, root cause diagnosis, and process optimization models based on newly added production data and control feedback data, continuously improving the model's multi-condition generalization and adaptation capabilities and data processing accuracy. At the same time, it is configured with a full-process data traceability algorithm to build a dedicated production data ledger for each batch of products.

[0050] The core iterative training steps of the incremental learning mechanism are as follows: 1. The newly added full-process operation data of the coating process each month is used as incremental samples. After cleaning and labeling, the samples are merged with the original standardized dataset. The incremental samples account for no less than 10% of the original dataset. 2. Freeze the training of the three core models of defect detection, root cause diagnosis and process optimization. Freeze the basic feature extraction layer of the model and only update the parameters of the output layer and fully connected layer. Set the learning rate to 1 / 10 of the initial training, i.e. 0.0001. 3. The accuracy of model detection, diagnosis and optimization is used as the verification metric. If the model accuracy improves by ≥2% after incremental training, the online deployment model parameters are updated. If the accuracy decreases, the incremental samples are discarded and the model is recalibrated. 4. Perform full-layer fine-tuning training on the model once per quarter to ensure the model is adapted to the flute type of the lamination production line. Changes in operating conditions such as membrane material and production speed; the iterative training data of each model are all linked to the production batch ledger, realizing full-process traceability of training data.

[0051] The production data ledger for a single batch of products is constructed in a structured table format. The core fields of the ledger include production batch number, production time, material attribute information, workstation identifiers for each process, key process parameter curves, defect detection results, root cause diagnosis data, early warning information, process optimization plans, feedback on control effects, and finished product quality inspection results. The specific execution steps of the data traceability algorithm are as follows: 1. Assign a unique production batch number to each batch of coated products and bind it to dual data tags; 2. Throughout the entire process of data collection and transmission, batch number and label information are carried. 3. The data closed-loop iteration module categorizes and stores all data from the same batch in a dedicated database, establishing a batch number index; 4. During traceability, the entire supply chain data can be retrieved by searching the index using the batch number. The system supports viewing, exporting, and tracing the data. The ledger data is permanently stored, meeting the industry requirements for production quality traceability.

[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent monitoring and fault diagnosis system for a bubble-free lamination process for corrugated cardboard boxes, characterized in that, It includes a data acquisition module, a spatiotemporal calibration module, a defect detection module, a root cause diagnosis module, a graded early warning module, a process optimization module, and a data closed-loop iteration module; The data acquisition module collects multi-dimensional, multi-source data from the entire coating process and configures dual data tags, transmitting the raw data to the spatiotemporal calibration module; The spatiotemporal calibration module performs spatiotemporal synchronization calibration and format standardization on the original multi-source data, and outputs the spatiotemporally synchronized standardized dataset to the defect detection module and the root cause diagnosis module after consistency verification. The defect detection module performs bubble defect detection on standardized visual image data and outputs structured detection data to the root cause diagnosis module and the graded early warning module. The root cause diagnosis module combines standardized operating condition parameter data with defect structured data to complete root cause localization and quantitative analysis, and outputs structured diagnostic data. The graded early warning module completes risk classification based on defect detection data and root cause diagnosis data, and generates early warning signals and control instructions. The process optimization module first verifies the solution through its built-in data processing logic, and then generates a process parameter adjustment plan based on the early warning data and root cause diagnosis data and sends it to the production line control system. The data closed-loop iteration module gathers the full-process operation data of each module and completes the adaptive iterative optimization of the system's core model.

2. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 1, characterized in that, The data acquisition module is an online real-time synchronous acquisition module, covering all process nodes from material pretreatment before lamination to roller lamination during lamination and cooling and shaping after lamination. The types of data collected include visual image data, equipment operation process parameter data, core component status data, workshop environmental condition data, and material attribute data. The dual data tags are uniquely encrypted dedicated workstation identifiers and hardware-level timestamps. The data acquisition module transmits the raw data with the configured tags to the spatiotemporal calibration module.

3. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 2, characterized in that, The spatiotemporal calibration module uses the synchronous pulse signal of the main shaft encoder of the coating production line to build a hardware-level timestamp alignment algorithm to complete the unified timing calibration of data with different sampling frequencies. Combining the spatial coordinate system of the coating production line and the workstation identification data label, it establishes a one-to-one correspondence between the location of visual defects, the corresponding process workstation, and the corresponding acquisition parameters. The spatiotemporal calibration module is equipped with a spatiotemporal misalignment correction algorithm and a data fusion verification mechanism to output a standardized dataset with spatiotemporal synchronization after eliminating invalid and misaligned data fragments.

4. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 3, characterized in that, The defect detection module incorporates a lightweight deep learning detection model specifically designed for coating scenarios. This lightweight deep learning detection model is optimized through multi-scale feature fusion and channel attention mechanisms, enabling it to identify all types of coating defects, including microbubbles, hidden interlayer bubbles, edge bubbles, and pinhole bubbles. The defect detection module performs frame-by-frame online detection on the visual images of the coating process, and the output structured detection data includes information on defect type, size, quantity, spatial location, and severity.

5. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 4, characterized in that, The root cause diagnosis module incorporates a causal structure learning model, which is trained using historical production data to construct a causal relationship network between bubble defects and various influencing parameters. The root cause diagnosis module uses a multi-parameter correlation decomposition analysis method to locate the core root cause parameters that induce bubble defects, and quantitatively determines the influence weight and risk level of the core root cause parameters. The output structured diagnostic data is transmitted to the graded early warning module, the process optimization module, and the data closed-loop iteration module.

6. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 5, characterized in that, The hierarchical early warning module has a built-in data-driven hierarchical early warning model. Based on three core indicators, such as defect severity, root cause risk weight, and defect propagation probability, a three-level early warning mechanism is established. The first level is a hidden risk warning for parameter anomalies and no visible defects. The second level is a local anomaly warning for minor defects and local data anomalies. The third level is a batch failure warning for severe defects and global data anomalies. Differentiated data management strategies are matched with early warning at each level, and early warning signals, graded data and disposal suggestions are pushed to the on-site management terminal and process optimization module simultaneously.

7. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 6, characterized in that, The process optimization module has a built-in adaptive optimization model for process parameters based on deep reinforcement learning. It takes bubble-free film coating as the core optimization objective and production efficiency and material loss control as data constraints. It combines real-time operating data to deduce and generate process parameter adjustment schemes. After the optimization scheme is verified by the built-in data processing logic, it is directly sent to the production line control system to realize intelligent process control.

8. The intelligent monitoring and fault diagnosis system for a bubble-free lamination process of corrugated cardboard boxes according to claim 7, characterized in that, The data closed-loop iteration module aggregates full-process operation data, including raw collected data, calibrated standardized data, defect detection data, root cause diagnosis data, early warning data, process optimization data, and control effect feedback data. This module builds a dedicated standardized dataset for the coating process, completing data cleaning, labeling, classification, storage, and compliance management. It incorporates an incremental learning mechanism and, based on newly added production data and control feedback data, performs adaptive incremental iterative training on the core models for defect detection, root cause diagnosis, and process optimization. Simultaneously, it configures a full-process data traceability algorithm to establish a dedicated production data ledger for each batch of products.

Images