Packaging box production line monitoring system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent manufacturing in the packaging industry, in particular to a packaging box production line monitoring system based on artificial intelligence, and aims to solve the problems that in the prior art, accurate feature extraction and classification cannot be carried out on packaging box images, multi-source information cannot be weighted and integrated by using an attention mechanism, and the detection accuracy is poor. And the accuracy of identifying the surface defects of the packaging box is improved. According to the invention, standardization, feature extraction and classification are carried out on packaging box images by using a convolutional neural network through a defect identification module, vibration, noise and temperature data collected from a die-cutting machine are processed and then fused with image features, multi-source information is weighted and integrated by using an attention mechanism, unified feature representation is generated, defect determination is completed by a full connection layer, and the defect identification accuracy is improved. And the accuracy and stability of the model in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology in the packaging industry, and more specifically, to an artificial intelligence-based packaging box production line monitoring system. Background Art

[0002] Traditional packaging box production lines have obvious technical defects in quality inspection and production management. It is difficult to strike a balance between efficiency and precision. Manual visual inspection is slow and prone to fatigue, resulting in a high missed detection rate. Mechanical sensors are restricted by the contact detection principle and cannot identify micron-level burst defects or complex structural deformations. The response lag problem is prominent, and equipment failure warnings rely on experience-based judgment, resulting in a long average downtime, affecting production continuity. Resources are seriously wasted, and excessive packaging leads to a high material loss rate. In addition, the traditional production scheduling system lacks dynamic optimization capabilities, making it difficult to flexibly adjust the production sequence according to real-time orders.

[0003] Patent application with reference publication number CN118605298A discloses an artificial intelligence-based packaging box production line monitoring system, including a durian packaging box status information acquisition module, a durian packaging box status information processing module, a production equipment status information acquisition module, a production equipment status information processing module, a production equipment status information processing module, a durian packaging box comprehensive analysis module, a durian packaging box production line abnormality monitoring module, a durian packaging box production line automation control module, and a user interface module; the monitoring range of the system of the present invention covers the entire packaging box production line, realizing comprehensive and accurate monitoring of the production line, and determining to improve product quality and production efficiency; it can dynamically adjust production parameters according to the real-time status of the production line and historical data analysis to maximize production efficiency and reduce costs, thereby enhancing the flexibility and intelligence level of production;

[0004] However, the above-mentioned reference patent dynamically optimizes production parameters through comprehensive and precise monitoring and intelligent control, determines to improve the efficiency and quality of the packaging production line, reduces energy consumption and costs, and promotes green and sustainable development, but cannot accurately extract and classify features of packaging box images, cannot use the attention mechanism to weightedly integrate multi-source information, and improve the accuracy of identifying surface defects of packaging boxes; at the same time, it cannot build a causal model based on time series and defect labels, cannot use image recognition to mark the defect moment, cannot extract variable lag features and fuse them into modeling, and cannot identify significant causal edges through Granger test, thereby reducing the accuracy and interpretability of root cause analysis.

[0005] To this end, we propose an artificial intelligence-based packaging production line monitoring system to address the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide an artificial intelligence-based packaging box production line monitoring system, which solves the problems that the existing technology cannot accurately extract and classify packaging box images, cannot use the attention mechanism to weightedly integrate multi-source information, and improve the accuracy of identifying surface defects of packaging boxes; at the same time, it cannot build a causal model based on time series and defect labels, cannot use image recognition to mark defect moments, cannot extract variable lag features and fuse them into modeling, and cannot identify significant causal edges through Granger test, thereby reducing the accuracy and interpretability of root cause analysis.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] An AI-based packaging production line monitoring system, applied to a packaging production management platform, includes:

[0009] The data acquisition and processing module is used to collect multi-source perception data from the packaging box production line and perform pre-processing operations on the collected multi-source perception data;

[0010] The defect recognition module uses a convolutional neural network to identify defects in packaging box images based on preprocessed multi-source sensor data. It uses the effective vibration value, ambient noise intensity, and ambient temperature of the die-cutting machine as auxiliary inputs and performs weighted fusion of multimodal features through an attention mechanism.

[0011] The fault root cause analysis module builds a structural causal model based on time series data and event labels, uses the Granger causality test method to extract the dynamic causal relationship between process parameters and equipment status, and traces abnormal paths and identifies key variable nodes for defect events;

[0012] The production optimization control module generates a scheduling plan through a genetic algorithm based on the process path, equipment status and energy consumption monitoring data, outputs it to the execution layer equipment, and updates the scheduling logic according to abnormal signals.

[0013] As a preferred embodiment of the present invention, the defect recognition module uses a convolutional neural network to identify defects in the packaging box image and combines the effective value of vibration, ambient noise intensity and ambient temperature during the operation of the die-cutting machine as auxiliary inputs, including the following process:

[0014] Obtain pre-processed multi-source sensing data, including the speed, pressure, and current of the die-cutting machine during operation, the surface image of the packaging box after die-cutting, and the effective value of vibration during the operation of the die-cutting machine;

[0015] Standardize the surface image of the die-cut packaging box;

[0016] Use convolutional layers to extract features from the preprocessed image to form a feature map. After applying the activation function, the size of the feature map is reduced through pooling operations to generate the final image feature representation.

[0017] The resulting image feature representation is fed into the fully connected layer, and the Softmax function is used to calculate the probability distribution of each possible defect category to determine whether there are defects in the image and their types.

[0018] Accelerometers are installed at key locations on the die-cutting machine to monitor the vibration signals generated by the machine in real time.

[0019] A microphone sensor is used to measure the noise intensity in the surrounding environment;

[0020] Use temperature sensors to monitor temperature changes in the die-cutting machine working area in real time;

[0021] The collected vibration effective value, ambient noise intensity and ambient temperature are standardized.

[0022] As a preferred embodiment of the present invention, the process of the defect recognition module performing weighted fusion of multimodal features through the attention mechanism includes the following steps:

[0023] The image feature vector extracted by the convolutional neural network and the standardized auxiliary data feature vectors of the three dimensions of vibration effective value, ambient noise intensity and ambient temperature are spliced ​​in a specific order to form a comprehensive feature vector;

[0024] The formed comprehensive feature vector is sent to a deep neural network structure containing several hidden layers for processing;

[0025] Based on the calculated image feature attention weights and auxiliary data feature attention weights, the image features and auxiliary features are weighted summed to generate a new fused feature representation;

[0026] The fused feature representation is used through a fully connected layer and combined with the Softmax function to make the final judgment on whether there is a defect and the specific defect type.

[0027] As a preferred embodiment of the present invention, the process of constructing a structural causal model based on time series data and event labels by the fault root cause analysis module includes:

[0028] During the industrial production process, sensors and control systems are used to monitor the equipment operating status and process parameters in real time, and the following types of data are recorded at a fixed sampling period: die-cutting pressure, equipment speed, motor current, vibration RMS value, bearing temperature, and ambient noise intensity;

[0029] A convolutional neural network is used to classify and judge the surface image of the packaging box. If there is a defect in the image, the corresponding moment is marked as the moment when the defect occurs, and a binary label sequence y is constructed. t ∈{0,1}, the label sequence is stored synchronously with the time series data;

[0030] The total number of collected process variables is defined as N, and the vector sequence of its components is:

[0031]

[0032] Each dimension represents the value of the i-th variable at time t;

[0033] In order to capture the dynamic dependency between variables, the lag characteristics of each variable are constructed, and the maximum lag order is defined as p. Then the feature vector of each variable at time t is composed of the values ​​of the current time and the previous p historical times. The lag characteristics of all variables are spliced ​​in order to obtain the comprehensive feature vector at time t.

[0034] The structural causal model is represented by a directed acyclic graph, denoted as G = (V, E), where:

[0035] Node set V = {X (1) ,X (2) ,…,X (N) ,Y} contains all variable nodes and defect event nodes;

[0036] The directed edge set E represents the direct causal relationship between variables;

[0037] If there is an edge X (i) →X (j) , indicating X (i) It's X (j) The direct dependent variable;

[0038] There is no closed loop path in the graph;

[0039] Identify graph structures from observed data that satisfy the following conditions:

[0040] Each edge in the graph represents a statistically determined causal relationship;

[0041] The predecessor node of the defect event node Y is its direct dependent variable;

[0042] All causal inference operations were performed according to established statistical testing criteria.

[0043] As a preferred embodiment of the present invention, the process of extracting the dynamic causal relationship between process parameters and equipment status by the fault root cause analysis module using the Granger causality test method includes:

[0044] Given two stationary time series and Define the following two regression models: an autoregressive model and an extended model with dependent variables;

[0045] The variable X (j) The historical data of are used to build the model containing only its own lag term and the model adding X (i) The extended model of the lag term calculates the residual sum of squares of the two models and constructs the F statistic. If the statistic is greater than the critical value under the specified degrees of freedom and significance level, it is determined that X (i) To X (j) It has Granger causality;

[0046] For any pair of variables, follow these steps to construct a causal edge:

[0047] Select the maximum lag order p, determined by the AIC or BIC criterion;

[0048] Construct AR model and VAR model respectively and calculate the residual sum of squares;

[0049] Calculate the F statistic to determine whether the significance test is passed;

[0050] If the test passes, add edge X (i) →X (j) ;

[0051] Perform the same test on each variable and the defect event label. If it is true, add an edge X (i) →Y;

[0052] Finally, a dynamic causal graph G = (V, E) supported by statistical reasoning is formed.

[0053] As a preferred embodiment of the present invention, the process of the fault root cause analysis module tracing the abnormal path and identifying the key variable nodes of the defect event includes:

[0054] In the causal graph, take the defect event as the target node and trace back all directed paths from any variable node to this node;

[0055] The variables appearing in the path are considered as direct or indirect antecedents of the defect event;

[0056] For a certain variable, if it appears in several paths and the total number of paths is known, the path coverage is defined as the ratio of the number of paths where the variable appears to the total number of paths;

[0057] To measure the information relevance between variables and defect events, the in-degree of each variable node in the directed graph is defined as the number of edges pointing to the node.

[0058] As a preferred embodiment of the present invention, the process of the production optimization control module generating a scheduling plan through a genetic algorithm based on the process path, equipment status and energy consumption monitoring data includes:

[0059] The system consists of several tasks, each of which must be completed in a fixed order on multiple devices. All tasks and devices have unique identifiers. The goal is to generate a scheduling plan that minimizes overall processing time, energy consumption, and exception penalties.

[0060] The objective function consists of three parts: maximum completion time, total energy consumption, and scheduling penalty. The three items are multiplied by the corresponding weight coefficients and then added together to form the final target value.

[0061] Each task must be completed in strict accordance with its prescribed order and on different devices;

[0062] If a device is unavailable, no tasks can be scheduled to execute on it.

[0063] The execution time and unit energy consumption of each task on each device are known parameters, based on which the total energy of the task on the device can be calculated;

[0064] Genetic algorithm is used to solve the scheduling problem. The specific steps are as follows:

[0065] T1: Generate an initial scheduling solution set according to preset rules;

[0066] T2: Calculate the corresponding target value for each scheduling scheme;

[0067] T3: Select individuals from the current population to enter the next generation based on fitness;

[0068] T4: Generate a new scheduling plan in a specific way while ensuring the execution order is legal;

[0069] T5: Adjust the individual task arrangements in the scheduling plan;

[0070] T6: directly retain the current optimal scheduling plan to the next generation;

[0071] Repeat the above steps until the termination condition is met, and finally output an optimal scheduling plan.

[0072] As a preferred embodiment of the present invention, the process of outputting the production optimization control module to the execution layer device and updating the scheduling logic according to the abnormal signal includes:

[0073] Convert the optimal scheduling plan into a standard format scheduling instruction list, each instruction contains the following information: task ID, equipment ID, start time, end time and expected energy consumption;

[0074] The real-time monitoring system continuously collects abnormal signals, including equipment status changes, excessive energy consumption, and task failure events;

[0075] When an abnormal signal is detected, the following actions are performed:

[0076] Analyze the exception type and determine the set of tasks affected;

[0077] Add dynamic penalty terms to fitness calculations;

[0078] Only the affected tasks will be rescheduled and optimized, while the unaffected tasks will remain in their original arrangements.

[0079] Only the updated instructions are reissued to the relevant devices.

[0080] Compared with the prior art, the advantages of the present invention are:

[0081] (1) In the present invention, a convolutional neural network is used to standardize, extract features, and classify packaging box images through a defect recognition module. The vibration, noise, and temperature data collected from the die-cutting machine are processed and fused with image features. The attention mechanism is used to weight and integrate multi-source information to generate a unified feature representation. The fully connected layer completes defect determination, thereby improving the accuracy and stability of the model in complex environments.

[0082] (2) In the present invention, a causal model is constructed based on time series and defect labels through the fault root cause analysis module, image recognition is used to mark the defect time, the target sequence is generated, the variable lag features are extracted and fused into the model, a directed acyclic graph is constructed to represent the causal relationship, and the significant causal edges are identified through the Granger test. After the defect occurs, the predecessor path is traced back, and the key factors are located by combining coverage, information relevance and in-degree indicators. The causal structure between variables is automatically derived, thereby improving the accuracy and interpretability of the root cause analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is a system block diagram of Embodiment 1 of the present invention;

[0084] Figure 2 This is a system block diagram of Embodiment 2 of the present invention;

[0085] Figure 3 This is a flow chart of the steps of using genetic algorithms to solve scheduling problems in the present invention. DETAILED DESCRIPTION

[0086] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work shall fall within the scope of protection of the present invention.

[0087] Example 1: Figure 1 As shown, the present invention proposes an artificial intelligence-based packaging box production line monitoring system, which is applied to a packaging box production management platform and includes:

[0088] The data acquisition and processing module is used to collect multi-source sensory data from the packaging production line and perform preprocessing operations on the collected multi-source sensory data. The preprocessing operations include noise removal, standardization, time stamp correction and unified format conversion;

[0089] The data acquisition and processing module collects and preprocesses multi-source sensory data from the packaging production line to improve data consistency and availability. Noise removal improves data quality, standardization achieves numerical comparability, timestamp correction ensures timing accuracy, and unified format conversion supports direct analysis by downstream systems. Overall, it provides a reliable data foundation for status monitoring, fault identification, and scheduling optimization.

[0090] The defect recognition module uses a convolutional neural network to identify defects in packaging box images based on preprocessed multi-source sensor data. It uses the effective vibration value, ambient noise intensity, and ambient temperature of the die-cutting machine as auxiliary inputs and performs weighted fusion of multimodal features through an attention mechanism.

[0091] The defect recognition module uses a convolutional neural network to identify defects in packaging box images and uses the effective vibration value, ambient noise intensity, and ambient temperature of the die-cutting machine as auxiliary inputs. The process includes:

[0092] Obtain pre-processed multi-source sensing data, including the speed, pressure, and current of the die-cutting machine during operation, the surface image of the packaging box after die-cutting, and the effective value of vibration during the operation of the die-cutting machine;

[0093] Normalize the surface image of the die-cut packaging box by dividing each pixel value by 255 and adjusting it to the range of [0,1].

[0094] Use convolutional layers to extract features from the preprocessed image to form a feature map. After applying the activation function, the size of the feature map is reduced through pooling operations to generate the final image feature representation.

[0095] The resulting image feature representation is fed into the fully connected layer, and the Softmax function is used to calculate the probability distribution of each possible defect category to determine whether there are defects in the image and their types.

[0096] Accelerometers are installed at key locations on the die-cutting machine to monitor the vibration signals generated by the machine in real time. The collected vibration signals are first filtered to remove noise interference, and then the root mean square algorithm is used to calculate the effective value of the vibration.

[0097] A microphone sensor is used to measure the noise intensity in the surrounding environment. After the collected sound signal is converted into an electrical signal, the spectrum characteristics are analyzed through fast Fourier transform to determine the frequency components of the main noise source. The equivalent continuous sound level is calculated based on the A-weighting network.

[0098] Use temperature sensors to monitor the temperature changes in the die-cutting machine's working area in real time. The temperature values ​​obtained directly reflect the degree of impact of environmental conditions on equipment performance.

[0099] The collected vibration RMS values, ambient noise intensity, and ambient temperature are standardized. Specifically, for each sensor data set, its mean and standard deviation are first calculated. Then, the mean of the set is subtracted from each observation value and then divided by its standard deviation to obtain a new data set with zero mean and unit variance.

[0100] The defect recognition module uses the attention mechanism to perform weighted fusion of multimodal features, including the following steps:

[0101] The image feature vector extracted by the convolutional neural network and the standardized auxiliary data feature vectors of the three dimensions of vibration effective value, ambient noise intensity and ambient temperature are spliced ​​in a specific order to form a comprehensive feature vector;

[0102] The resulting comprehensive feature vector is fed into a deep neural network structure containing several hidden layers for processing. The specific processing steps are as follows:

[0103] S1: For the image feature part, linear transformation and nonlinear activation operations are performed, and then the Softmax function is applied to obtain the attention weight of the image feature;

[0104] S2: For the auxiliary data feature part, linear transformation and nonlinear activation operations are performed, and then the Softmax function is applied to obtain the attention weight of the auxiliary data feature;

[0105] According to the calculated image feature attention weight and auxiliary data feature attention weight, the image feature and the auxiliary feature are weighted and summed to generate a new fused feature representation. Specifically, the image feature is multiplied by its corresponding attention weight, the auxiliary data feature is also multiplied by its corresponding attention weight, and then the two are added to obtain the fused feature representation;

[0106] The fused feature representation is then passed through a fully connected layer and combined with the Softmax function to determine whether a defect exists and its specific type. This design fully leverages the advantages of multi-source heterogeneous information and improves the robustness and generalization capabilities of the model.

[0107] The defect recognition module uses a convolutional neural network to standardize, extract features, and classify the packaging box image. At the same time, the vibration effective value, ambient noise intensity, and temperature data of the die-cutting machine are collected, filtered, spectrally analyzed, and standardized, and then spliced ​​with the image features into a comprehensive feature vector. The attention weights of each part are calculated through a deep neural network and weighted fusion is performed to generate a unified feature representation. Finally, the fully connected layer completes the determination of the defect type. This method integrates the image and equipment operation status information, improving the recognition accuracy and stability of the model in complex environments.

[0108] The fault root cause analysis module builds a structural causal model based on time series data and event labels, uses the Granger causality test method to extract the dynamic causal relationship between process parameters and equipment status, and traces abnormal paths and identifies key variable nodes for defect events;

[0109] The process of constructing a structural causal model based on time series data and event labels in the fault root cause analysis module includes:

[0110] During the industrial production process, sensors and control systems are used to monitor the equipment operating status and process parameters in real time, and the following types of data are recorded at a fixed sampling period: die-cutting pressure, equipment speed, motor current, vibration RMS value, bearing temperature, and ambient noise intensity;

[0111] All data is transmitted to a central database via industrial communication protocols and stored aligned by timestamp to form a multidimensional time series dataset;

[0112] A convolutional neural network is used to classify and judge the surface image of the packaging box. If there is a defect in the image, the corresponding moment is marked as the moment when the defect occurs, and a binary label sequence y is constructed. t ∈{0,1}, where:

[0113]

[0114] The label sequence is stored synchronously with the time series data and serves as the target variable for subsequent structural causal modeling;

[0115] The total number of collected process variables is defined as N, and the vector sequence of its components is:

[0116]

[0117] Each dimension represents the value of the i-th variable at time t, and the sampling period remains constant;

[0118] To capture the dynamic dependency between variables, the lag characteristics of each variable are constructed, and the maximum lag order is defined as p. Then, the feature vector of each variable at time t is composed of the values ​​at the current time and the values ​​at the previous p historical times. The lag characteristics of all variables are sequentially spliced ​​to obtain the comprehensive feature vector at time t, which is used for subsequent causal modeling.

[0119] The structural causal model is represented by a directed acyclic graph, denoted as G = (V, E), where:

[0120] Node set V = {X (1) ,X (2) ,…,X (N) ,Y} contains all variable nodes and defect event nodes;

[0121] The directed edge set E represents the direct causal relationship between variables;

[0122] If there is an edge X (i) →X (j) , indicating X (i) It's X (j) The direct dependent variable;

[0123] There is no closed loop path in the graph;

[0124] Identify graph structures from observed data that satisfy the following conditions:

[0125] Each edge in the graph represents a statistically determined causal relationship;

[0126] The predecessor node of the defect event node Y is its direct dependent variable;

[0127] All causal inference operations are performed according to established statistical testing criteria;

[0128] The fault root cause analysis module uses the Granger causality test method to extract the dynamic causal relationship between process parameters and equipment status. The process includes:

[0129] Given two stationary time series and Define the following two regression models:

[0130] Autoregressive Model: in represents the value of the jth variable at time t, represents the value of the jth variable k time units ago, ε t Represents the residual term of the model, which represents the part that cannot be explained by the model, a k represents the k-th order regression coefficient in the autoregressive model, and p represents the maximum lag order, that is, the length of historical time considered by the model;

[0131] Expanded model with dependent variables: in represents the value of the i-th variable k time units ago, b k represents the regression coefficient of variable i introduced in the kth order in the extended model, ε t ′ represents the residual term of the model, which represents the part that cannot be explained by the model;

[0132] The variable X (j) The historical data of are used to build the model containing only its own lag term and the model adding X (i) The extended model of the lag term calculates the residual sum of squares of the two models and constructs the F statistic. If the statistic is greater than the critical value under the specified degrees of freedom and significance level, it is determined that X (i) To X (j) It has Granger causality;

[0133] For any pair of variables, follow these steps to construct a causal edge:

[0134] Select the maximum lag order p, determined by the AIC or BIC criterion;

[0135] Construct AR model and VAR model respectively and calculate the residual sum of squares;

[0136] Calculate the F statistic to determine whether the significance test is passed;

[0137] If the test passes, add edge X (i) →X (j) ;

[0138] Perform the same test on each variable and the defect event label. If it is true, add an edge X (i) →Y;

[0139] Finally, a dynamic causal graph G = (V, E) supported by statistical reasoning is formed;

[0140] The fault root cause analysis module traces the abnormal path of the defect event and identifies the key variable nodes, including the following:

[0141] In the causal graph, with the defect event as the target node, all directed paths from any variable node to the node are traced. All paths must satisfy directional consistency. Path search can be achieved by reverse traversal of the graph or topological sorting.

[0142] The variables appearing in the path are considered as direct or indirect antecedents of the defect event;

[0143] For a variable, if it appears in several paths and the total number of paths is known, the path coverage is defined as the ratio of the number of paths in which the variable appears to the total number of paths. The higher the value, the more critical the variable is.

[0144] Measures the information relevance between the variable and the defect event. If the value is significantly greater than that of other variables, it means that the variable has a stronger information relevance to the defect event.

[0145] The in-degree of each variable node in a directed graph is defined as the number of edges pointing to the node. Nodes with high in-degree are usually located on multiple paths and are "mediating variables" in the structure.

[0146] The fault root cause analysis module constructs a structural causal model based on time series data and defect event labels. Sensors collect multidimensional variable data such as die-cutting pressure, equipment speed, motor current, vibration RMS, bearing temperature, and ambient noise intensity, and store them aligned by timestamp. A convolutional neural network is used to identify image defects, mark the time when the defects occur, and form a target variable sequence. A lag feature is constructed for each variable and spliced ​​into a comprehensive feature vector for modeling dynamic dependencies. The structural causal model represents the direct causal relationship between variables using a directed acyclic graph. The Granger causality test method is used to identify statistically significant causal edges and construct a dynamic causal graph. After a defect event occurs, all predecessor paths are traced, and variable coverage, information relevance, and in-degree centrality indicators are statistically analyzed to identify key influencing factors. This method automatically derives the causal structure between variables from the data, supports systematic attribution analysis of defect events, and improves the accuracy and interpretability of root cause location.

[0147] Embodiment 2: The technical solution of this embodiment of the present invention differs from that of Embodiment 1 in that:

[0148] like Figure 2 and Figure 3 As shown in the figure, the production optimization control module generates a scheduling plan based on the process path, equipment status and energy consumption monitoring data through a genetic algorithm, outputs it to the execution layer equipment, and updates the scheduling logic according to abnormal signals;

[0149] The production optimization control module generates a scheduling plan based on process paths, equipment status, and energy consumption monitoring data through a genetic algorithm. The process includes:

[0150] The system consists of several tasks, each of which must be completed in a fixed order on multiple devices. All tasks and devices have unique identifiers. The goal is to generate a scheduling plan that minimizes overall processing time, energy consumption, and exception penalties.

[0151] The objective function consists of three parts:

[0152] Maximum completion time: refers to the latest time point among all task completion times;

[0153] Total energy consumption: refers to the total energy consumed by all tasks when they are executed on the devices to which they are assigned;

[0154] Scheduling penalty: used to represent the additional cost caused by device conflicts or abnormal conditions;

[0155] The three items are multiplied by the corresponding weight coefficients and then added together to form the final target value;

[0156] Each task must be completed in strict accordance with its prescribed order and on different devices;

[0157] If a device is unavailable, no tasks can be scheduled to execute on it.

[0158] The execution time and unit energy consumption of each task on each device are known parameters, based on which the total energy of the task on the device can be calculated;

[0159] The scheduling problem is solved using a genetic algorithm. The chromosome consists of two parts: the first part represents the execution order of tasks, and the second part represents the equipment assigned to each task and its execution time period. The specific steps are as follows:

[0160] T1: Generate an initial scheduling solution set according to preset rules;

[0161] T2: Calculate the corresponding target value for each scheduling scheme;

[0162] T3: Select individuals from the current population to enter the next generation based on fitness;

[0163] T4: Generate a new scheduling plan in a specific way while ensuring the execution order is legal;

[0164] T5: Adjust the individual task arrangements in the scheduling plan;

[0165] T6: directly retain the current optimal scheduling plan to the next generation;

[0166] Repeat the above steps until the termination condition is met, and finally output an optimal scheduling plan;

[0167] The process of outputting the production optimization control module to the execution layer equipment and updating the scheduling logic based on the abnormal signal includes:

[0168] Convert the optimal scheduling plan into a standard format scheduling instruction list, each instruction contains the following information: task ID, equipment ID, start time, end time and expected energy consumption;

[0169] All instructions are arranged in chronological order and sent to the control system through the industrial communication protocol. After receiving the instructions, the control system calls the interface to load the task to the device controller. Each device only processes the instructions that match its identification and executes them in chronological order.

[0170] The real-time monitoring system continuously collects abnormal signals, including equipment status changes, energy consumption exceeding standards, and task failure events. Each abnormal signal contains the following information: equipment identification, occurrence time, abnormality type, and severity level;

[0171] When an abnormal signal is detected, the following actions are performed:

[0172] Analyze the exception type and determine the set of tasks affected;

[0173] Add dynamic penalty terms to fitness calculations to reflect the impact of anomalies;

[0174] Only the affected tasks will be rescheduled and optimized, while the unaffected tasks will remain in their original arrangements.

[0175] Only updated instructions are reissued to relevant devices to avoid disturbances to the entire system;

[0176] The production optimization control module uses a genetic algorithm to generate a scheduling plan based on process paths, equipment status, and energy consumption data, and sends it to the execution-layer equipment. Through iterative optimization through operations such as initializing the population, calculating fitness, selection, crossover, and mutation, the optimal solution that takes into account completion time, energy consumption, and abnormality penalties is obtained. Scheduling instructions are encapsulated in a standard format and transmitted to the equipment controller via an industrial communication protocol to ensure that tasks are executed on time. The system collects abnormal signals in real time, dynamically adjusts the fitness function after identifying affected tasks, and only reschedules local tasks and updates related instructions to avoid affecting the overall process. This method realizes adaptive adjustment of the scheduling plan and improves the stability and response efficiency of the production system.

[0177] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solutions and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the scope of protection of the present invention.

Claims

1. An artificial intelligence-based packaging production line monitoring system, applied to a packaging production management platform, characterized by: include: The data acquisition and processing module is used to collect multi-source perception data from the packaging box production line and perform pre-processing operations on the collected multi-source perception data; The defect recognition module uses a convolutional neural network to identify defects in packaging box images based on preprocessed multi-source sensor data. It uses the effective vibration value, ambient noise intensity, and ambient temperature of the die-cutting machine as auxiliary inputs and performs weighted fusion of multimodal features through an attention mechanism. The fault root cause analysis module builds a structural causal model based on time series data and event labels, uses the Granger causality test method to extract the dynamic causal relationship between process parameters and equipment status, and traces abnormal paths and identifies key variable nodes for defect events; The production optimization control module generates a scheduling plan through a genetic algorithm based on the process path, equipment status and energy consumption monitoring data, outputs it to the execution layer equipment, and updates the scheduling logic according to abnormal signals.

2. The artificial intelligence-based packaging production line monitoring system according to claim 1, characterized in that: The defect recognition module uses a convolutional neural network to identify defects in the packaging box image and combines the effective value of vibration, ambient noise intensity and ambient temperature during the operation of the die-cutting machine as auxiliary inputs. The process includes: Obtain pre-processed multi-source sensing data, including the speed, pressure, and current of the die-cutting machine during operation, the surface image of the packaging box after die-cutting, and the effective value of vibration during the operation of the die-cutting machine; Standardize the surface image of the die-cut packaging box; Use convolutional layers to extract features from the preprocessed image to form a feature map. After applying the activation function, the size of the feature map is reduced through pooling operations to generate the final image feature representation. The resulting image feature representation is fed into the fully connected layer, and the Softmax function is used to calculate the probability distribution of each possible defect category to determine whether there are defects in the image and their types. Accelerometers are installed at key locations on the die-cutting machine to monitor the vibration signals generated by the machine in real time. A microphone sensor is used to measure the noise intensity in the surrounding environment; Use temperature sensors to monitor temperature changes in the die-cutting machine working area in real time; The collected vibration effective value, ambient noise intensity and ambient temperature are standardized.

3. The artificial intelligence-based packaging production line monitoring system according to claim 2, characterized in that: The process of the defect recognition module weightedly fusing multimodal features through the attention mechanism includes: The image feature vector extracted by the convolutional neural network and the standardized auxiliary data feature vectors of the three dimensions of vibration effective value, ambient noise intensity and ambient temperature are spliced ​​in a specific order to form a comprehensive feature vector; The formed comprehensive feature vector is sent to a deep neural network structure containing several hidden layers for processing; Based on the calculated image feature attention weights and auxiliary data feature attention weights, the image features and auxiliary features are weighted summed to generate a new fused feature representation; The fused feature representation is used through a fully connected layer and combined with the Softmax function to make the final judgment on whether there is a defect and the specific defect type.

4. The artificial intelligence-based packaging production line monitoring system according to claim 1, characterized in that: The process of constructing a structural causal model based on time series data and event labels by the fault root cause analysis module includes: During the industrial production process, sensors and control systems are used to monitor the equipment operating status and process parameters in real time, and the following types of data are recorded at a fixed sampling period: die-cutting pressure, equipment speed, motor current, vibration RMS value, bearing temperature, and ambient noise intensity; A convolutional neural network is used to classify and judge the surface image of the packaging box. If there is a defect in the image, the corresponding moment is marked as the moment when the defect occurs, and a binary label sequence y is constructed. t ∈{0,1}, the label sequence is stored synchronously with the time series data; The total number of collected process variables is defined as N, and the vector sequence of its components is: Each dimension represents the value of the i-th variable at time t; In order to capture the dynamic dependency between variables, the lag characteristics of each variable are constructed, and the maximum lag order is defined as p. Then the feature vector of each variable at time t is composed of the values ​​of the current time and the previous p historical times. The lag characteristics of all variables are spliced ​​in order to obtain the comprehensive feature vector at time t. The structural causal model is represented by a directed acyclic graph, denoted as G = (V, E), where: Node set V = {X (1) ,X (2) ,…,X (N) ,Y} contains all variable nodes and defect event nodes; The directed edge set E represents the direct causal relationship between variables; If there is an edge X (i) →X (j) , indicating X (i) It's X (j) The direct dependent variable; There is no closed loop path in the graph; Identify graph structures from observed data that satisfy the following conditions: Each edge in the graph represents a statistically determined causal relationship; The predecessor node of the defect event node Y is its direct dependent variable; All causal inference operations were performed according to established statistical testing criteria.

5. The artificial intelligence-based packaging production line monitoring system according to claim 4, characterized in that: The process of extracting the dynamic causal relationship between process parameters and equipment status using the Granger causality test method in the fault root cause analysis module includes: Given two stationary time series and Define the following two regression models: an autoregressive model and an extended model with dependent variables; The variable X (j) The historical data of are used to build the model containing only its own lag term and adding X (i) The extended model of the lag term calculates the residual sum of squares of the two models and constructs the F statistic. If the statistic is greater than the critical value under the specified degrees of freedom and significance level, it is determined that X (i) To X (j) It has Granger causality; For any pair of variables, follow these steps to construct a causal edge: Select the maximum lag order p, determined by the AIC or BIC criterion; Construct AR model and VAR model respectively and calculate the residual sum of squares; Calculate the F statistic to determine whether the significance test is passed; If the test passes, add edge X (i) →X (j) ; Perform the same test on each variable and the defect event label. If it is true, add an edge X (i) →Y; Finally, a dynamic causal graph G = (V, E) supported by statistical reasoning is formed.

6. The artificial intelligence-based packaging production line monitoring system according to claim 5, characterized in that: The process of the fault root cause analysis module tracing the abnormal path of the defect event and identifying the key variable nodes includes: In the causal graph, take the defect event as the target node and trace back all directed paths from any variable node to this node; The variables appearing in the path are considered as direct or indirect antecedents of the defect event; For a certain variable, if it appears in several paths and the total number of paths is known, the path coverage is defined as the ratio of the number of paths where the variable appears to the total number of paths; To measure the information relevance between variables and defect events, the in-degree of each variable node in the directed graph is defined as the number of edges pointing to the node.

7. The artificial intelligence-based packaging production line monitoring system according to claim 1, characterized in that: The process of the production optimization control module generating a scheduling plan based on the process path, equipment status and energy consumption monitoring data through a genetic algorithm includes: The system consists of several tasks, each of which must be completed in a fixed order on multiple devices. All tasks and devices have unique identifiers. The goal is to generate a scheduling plan that minimizes overall processing time, energy consumption, and exception penalties. The objective function consists of three parts: maximum completion time, total energy consumption, and scheduling penalty. The three items are multiplied by the corresponding weight coefficients and then added together to form the final target value. Each task must be completed in strict accordance with its prescribed order and on different devices; If a device is unavailable, no tasks can be scheduled to execute on it. The execution time and unit energy consumption of each task on each device are known parameters, based on which the total energy of the task on the device can be calculated; Genetic algorithm is used to solve the scheduling problem. The specific steps are as follows: T1: Generate an initial scheduling solution set according to preset rules; T2: Calculate the corresponding target value for each scheduling scheme; T3: Select individuals from the current population to enter the next generation based on fitness; T4: Generate a new scheduling plan in a specific way while ensuring the execution order is legal; T5: Adjust the individual task arrangements in the scheduling plan; T6: directly retain the current optimal scheduling plan to the next generation; Repeat the above steps until the termination condition is met, and finally output an optimal scheduling plan.

8. The artificial intelligence-based packaging production line monitoring system according to claim 7, characterized in that: The process of outputting the production optimization control module to the execution layer device and updating the scheduling logic according to the abnormal signal includes: Convert the optimal scheduling plan into a standard format scheduling instruction list, each instruction contains the following information: task ID, equipment ID, start time, end time and expected energy consumption; The real-time monitoring system continuously collects abnormal signals, including equipment status changes, excessive energy consumption, and task failure events; When an abnormal signal is detected, the following actions are performed: Analyze the exception type and determine the set of tasks affected; Add dynamic penalty terms to fitness calculations; Only the affected tasks will be rescheduled and optimized, while the unaffected tasks will remain in their original arrangements. Only the updated instructions are reissued to the relevant devices.

Images