A dynamic visual monitoring under packaging production line early warning method and system

Abstract

The application belongs to the technical field of deep learning, and discloses a packaging production line early warning method and system under dynamic visual monitoring, which comprises continuous image acquisition of the packaging production line, image sequence acquisition of the packaging production line; semantic feature information representing the packaging process state is extracted from the image sequence, a multi-dimensional tensor flow data structure is constructed based on the semantic feature information, and a packaging process semantic atlas is generated; the semantic behavior curvature theory is introduced, the packaging process semantic atlas is subjected to curvature analysis, the semantic evolution curvature change rate of different time sequences is calculated, and a sliding window dynamic curvature detection mechanism is used to identify the existing abnormal bending area and generate semantic bending abnormal information; the deficiencies of traditional video monitoring methods in process complexity, abnormal fine granularity identification and response timeliness are solved, and a more intelligent, accurate and adaptive early warning decision-making closed loop is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, more particularly, the present application relates to a dynamic visual monitoring packaging production line early warning method and system. BACKGROUND

[0002] The patent with the patent publication number CN120145272A discloses an intelligent packaging production line control system and method, aiming to realize comprehensive monitoring and fault early warning of the intelligent packaging production line; including: collecting power parameter data of electrical components; the power parameter data includes current, voltage, energy utilization ratio data and vibration signal of the electrical components; the power parameter data is preprocessed to form an electrical feature data set; a fault state prediction model is trained according to the electrical feature data set, and a fault probability value of the electrical component is predicted based on the fault state prediction model; according to the fault probability value of the electrical component, it is judged whether the intelligent packaging production line has a fault; when it is detected that the intelligent packaging production line has a fault, the electrical component that appears the fault is analyzed, and targeted control measures are taken to ensure efficient fault identification and control in various special environments.

[0003] The existing packaging production line early warning method and system mainly have the following problems:

[0004] Most of the existing methods use fixed parameter configuration, and the sensitivity to packaging process state changes cannot be dynamically adjusted according to the scene, which leads to excessive response of the system to normal fluctuations or insensitivity to potential instability in actual operation, affecting the stability and reliability of the system. The existing system relies on single visual features such as image recognition or action detection, and triggers an early warning as soon as a change is recognized. However, in actual packaging processes, many disturbances belong to the category of normal fluctuations, such as mechanical micro-shaking, light fluctuation, etc. Although the above disturbances can cause visual differences, they do not mean that the system is at risk. Traditional methods cannot effectively distinguish between normal disturbances and potential abnormal evolution, and often treat all changes uniformly, resulting in high false positive rate and low system credibility; these disturbances have the characteristics of short time, frequent and low risk, but traditional methods are difficult to effectively identify their differences, which easily causes frequent false positives and operational interference.

[0005] The existing mechanism focuses on the state of the current frame, lacks trend learning or suppression of the disturbance history, causes the system to overreact when facing continuous high-frequency disturbances, lacks self-inhibition ability, and reduces the robustness of the early warning system. Many early warning systems regard the change of the package state as an isolated event, and rely on the state of a frame or a short time to determine whether there is an anomaly, lacking modeling and utilization of the state evolution trend and causal trajectory. This can cause the system to fail to identify the gradual instability trend, such as the gradual tilting of the package until it falls; fail to capture the cascading failure triggered by multiple slight deviations, and ignore the gradual abnormal behavior in the production process, such as position drift and motion deceleration. In actual production, the packaging process has dynamic change scenes such as task switching, beat change, equipment aging, and manual intervention. In the existing technology, the early warning strategy is mostly based on the current state to make an immediate judgment, rather than combining the historical intervention effect and scene evolution record for intelligent backtracking and strategy optimization. This causes the response strategy to fail to automatically optimize or evolve, and still needs to be recalculated in repeated scenes, making it difficult to improve processing efficiency and accuracy;

[0006] The traditional packaging production line visual monitoring system mostly uses a fixed frequency and uniformly collects all regions, resulting in a large amount of invalid image data and serious resource waste. The existing technology cannot timely perceive the dynamic changes of different key stations and regions in the packaging process, and cannot quickly identify process abnormalities or state mutations, affecting the timeliness and accuracy of abnormal early warning. The traditional method usually assigns static or empirical weights to the monitoring regions, resulting in insufficient sampling frequency of key change regions and unsatisfactory abnormal monitoring effect. The system collects a large amount of redundant images, increasing the storage and computing burden, and reducing the real-time performance and response speed of the overall monitoring system. Noise filtering, brightness normalization, and frame alignment processing of the image sequence are not in place, affecting the accuracy of subsequent semantic feature extraction, and further affecting the accurate identification of the packaging process state and abnormal bending detection.

[0007] In view of this, the present application proposes a packaging production line early warning method under dynamic visual monitoring to solve the above problems. SUMMARY

[0008] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a packaging production line early warning method under dynamic visual monitoring, comprising:

[0009] S1, continuously image collecting the packaging production line to obtain an image sequence of the packaging production line; extracting semantic feature information representing the packaging process state from the image sequence, and constructing a multi-dimensional tensor flow data structure based on the semantic feature information to generate a packaging process semantic atlas;

[0010] S2, introduce semantic behavior curvature theory, perform curvature analysis on the packaging process semantic atlas, calculate semantic evolution curvature change rate of different time sequences, and use a sliding window dynamic curvature detection mechanism to identify abnormal bending areas and generate semantic bending abnormal information;

[0011] S3, construct an endogenous disturbance response function based on the current running state vector of the nodes in the packaging process semantic atlas, calculate the response sensitivity of the endogenous disturbance response function in the preset disturbance direction, and output an endogenous disturbance energy index representing the potential instability trend of the current packaging state;

[0012] S4, synchronously fuse and analyze the semantic bending abnormal information and the endogenous disturbance energy index, determine the risk level of the packaging production line according to the fusion result, and generate corresponding graded warning signals;

[0013] S5, trigger the corresponding response strategy according to the graded warning signal, and generate a semantic disturbance memory atlas based on the response strategy; when a similar semantic bending behavior or energy disturbance trend is identified, the atlas similarity is used to preferentially match the optimal response strategy, supporting the evolution and update of the warning mechanism.

[0014] Preferably, the image sequence acquisition method comprises:

[0015] Industrial vision cameras are deployed at different key workstations of the packaging production line to continuously capture images of the packaging operation process; the industrial vision cameras are set by fixed installation or mechanical arm tracking, and a time stamp is added to each frame of image;

[0016] The key workstations include a feeding workstation, a filling workstation, a correction workstation, a packaging workstation, a labeling workstation, a code spraying workstation, a viewing workstation, a boxing workstation, and a sorting workstation; the image sequence is captured in a continuous frame capture mode, and the capture frequency is set according to the packaging line beat; a ring buffer structure is used to cache video image data;

[0017] An image difference-based region triggering mechanism is introduced, the image is divided into m regions, the pixel change ratio of the image difference graph is calculated for each region, and when the pixel change ratio of any region exceeds a preset pixel change ratio threshold, it is determined that the region is in a process change state, triggering dynamic enhanced sampling of the image frame where the process change occurs; and the image sequence is subjected to noise filtering, brightness normalization, and frame alignment operations to generate a standardized image sequence.

[0018] Preferably, the packaging process semantic atlas acquisition method comprises:

[0019] The semantic elements involved in the packaging process in the image sequence are detected and classified by using a semantic recognition model based on a deep neural network; the semantic recognition model is any one of a deep convolutional neural network or a target detection network model; the semantic elements include a workstation type, an operation object, an action state, a position coordinate, and time information;

[0020] The recognition results are encoded into fixed-length semantic state vectors by using One-hot encoding, numerical normalization, and position encoding methods, and are organized in the order of image frame time and workstation space to construct a multi-dimensional tensor flow data structure containing time, space, and semantic dimensions;

[0021] Each fixed-length semantic state vector in the multi-dimensional tensor flow data structure is mapped to a node in a semantic graph, and an edge connection between the nodes is established according to the time evolution relationship and the operation process causal relationship to form a packaging process semantic graph reflecting the semantic structure and evolution path of the packaging production process.

[0022] Preferably, the method for obtaining the semantic evolution curvature change rate of different time sequences comprises:

[0023] In the packaging process semantic graph, a continuous semantic state node sequence of the same object or the same workstation is selected along the time evolution relationship edge; based on the continuous semantic state node sequence, a semantic state vector trajectory is extracted, and the semantic state vector trajectory is mapped to a semantic state space to form a semantic evolution path;

[0024] The curvature of the semantic evolution path is calculated by using a three-point discrete curvature estimation, a difference operation is performed on the semantic state vectors corresponding to the continuous three semantic state nodes, the curvature of the middle node relative to the front and rear nodes of the middle node is calculated, and the change speed of the curvature with time is counted within a preset time window to obtain the semantic evolution curvature change rate of different time sequences.

[0025] Preferably, the method for obtaining the semantic bending abnormal information comprises:

[0026] The semantic evolution curvature change rates of the time sequences are arranged in time sequence to form a semantic evolution curvature change rate sequence; a preset fixed-length sliding time window is slid on the semantic evolution curvature change rate sequence at a preset step to extract a curvature change rate segment corresponding to each sliding window;

[0027] For each sliding window segment, a statistical characteristic index of the semantic evolution curvature change rate in the sliding window is calculated, the statistical characteristic index includes an average curvature change rate, a maximum curvature change rate, and a standard deviation of the curvature change rate, and the statistical characteristic index is compared with a preset abnormal statistical characteristic index determination threshold;

[0028] When the statistical feature index exceeds the preset statistical feature index threshold, it is determined that the time period corresponding to the current window has an abnormal bending trend of semantic state evolution; adjacent or overlapping abnormal windows are merged to form an abnormal interval, and the start and end time, the associated semantic state node set and the corresponding abnormal intensity index of the abnormal interval are output to form semantic bending abnormal information.

[0029] Preferably, the method for constructing the endogenous disturbance response function comprises:

[0030] A preset target node is selected from the packaging process semantic graph, and the preset target node corresponds to the running state of any key station of the packaging production line at the current time; a current running state vector of the preset target node is extracted, and the running state vector includes a station type code, operation object information, action parameters, spatial position information and a time label;

[0031] A set of disturbance direction vectors for representing potential changes in the packaging process are preset, and the disturbance direction vectors include process parameter disturbance vectors, position disturbance vectors and semantic category disturbance vectors;

[0032] For each disturbance direction, the response sensitivity of the running state vector in the disturbance direction is calculated based on the linear projection relationship between the running state vector of the current node and the disturbance direction and the state disturbance gradient, the endogenous disturbance response function is constructed, and the disturbance response weight coefficient in the endogenous disturbance response function is adaptively adjusted.

[0033] Preferably, the method for obtaining the endogenous disturbance energy index comprises:

[0034] Based on the constructed endogenous disturbance response function, the response sensitivity of the preset target node in different disturbance direction vectors is calculated, and the response degree of the current running state vector to the semantic state change in each disturbance direction is quantified by the endogenous disturbance response function;

[0035] The quantification process includes multiplying the linear projection result of the current running state vector on the disturbance direction vector by the state disturbance gradient in the disturbance direction to obtain the response intensity value of each disturbance direction; the response intensity values in the disturbance directions are normalized to obtain the overall disturbance response sensitivity index of the current packaging state as the endogenous disturbance energy index representing the potential instability trend of the current packaging state.

[0036] Preferably, the method for determining the risk level of the packaging production line and generating a corresponding graded warning signal comprises:

[0037] The semantic bending abnormal information and the endogenous disturbance energy index are synchronized and time-aligned, corresponding matching and normalization processing are performed in a unified time window; a fusion function is constructed, the normalized semantic bending abnormal information and the endogenous disturbance energy index are weighted and fused according to a preset weight to generate a comprehensive risk index;

[0038] According to the comprehensive risk index, in combination with a preset risk threshold, the current state of the packaging production line is divided into different risk levels, and the risk levels include a normal level, a warning level, a serious level and a critical level; according to the determined risk level, a corresponding graded early warning signal is generated, and the graded early warning signal contains risk level information, an abnormal time range, an abnormal type and an identification of a related station.

[0039] Preferably, the method of utilizing the atlas similarity to preferentially match the optimal response strategy comprises:

[0040] According to the risk level corresponding to the graded early warning signal, a matched response strategy is triggered, and the response strategy includes production rhythm adjustment, key station suspension, manual intervention reminder and process correction instruction execution; during the execution of the response strategy, a response event in a current time period is recorded; the response event includes a packaging process semantic atlas structure subgraph in a current time window, corresponding semantic bending abnormal information, an endogenous disturbance energy index, a risk level and a triggered matched response strategy;

[0041] The response event is structured and organized to construct a semantic disturbance memory atlas, and the association relationship between historical early warning and response events is stored; a substructure in the current semantic disturbance memory atlas is identified, and based on a subgraph isomorphism detection algorithm, the substructure is compared with a subgraph recorded in the semantic disturbance memory atlas in structure and similarity calculation is performed;

[0042] When it is detected that the similarity between the current semantic bending behavior and the semantic bending behavior of the recorded subgraph in the historical semantic disturbance memory atlas is greater than a preset similarity threshold, the optimal response strategy corresponding to the historical semantic disturbance memory atlas is immediately preferentially called to perform early warning response strategy calling and dynamic optimization updating.

[0043] A packaging production line early warning system under dynamic visual monitoring, comprising:

[0044] A multi-dimensional visual semantic module performs continuous image acquisition on the packaging production line to obtain an image sequence of the packaging production line; semantic feature information representing a packaging process state is extracted from the image sequence, and a multi-dimensional tensor flow data structure is constructed based on the semantic feature information to generate a packaging process semantic atlas;

[0045] A semantic bending degree analysis module is introduced to introduce a semantic behavior curvature theory, to perform curvature analysis on the packaging process semantic graph, to calculate semantic evolution curvature change rates of different time sequences, and to utilize a sliding window dynamic curvature detection mechanism to identify abnormal bending areas, and to generate semantic bending abnormal information;

[0046] An endogenous disturbance energy calculation module is based on a current running state vector of nodes in the packaging process semantic graph to construct an endogenous disturbance response function, to calculate a response sensitivity of the endogenous disturbance response function in a preset disturbance direction, and to output an endogenous disturbance energy index representing a potential instability trend of a current packaging state;

[0047] A collaborative risk judgment module is used to perform synchronous fusion analysis on the semantic bending abnormal information and the endogenous disturbance energy index, to judge a risk level of the packaging production line according to a fusion result, and to generate a corresponding graded early warning signal;

[0048] An early warning response evolution module is used to trigger a corresponding response strategy according to the graded early warning signal, and to generate a semantic disturbance memory graph based on the response strategy; when a similar semantic bending behavior or energy disturbance trend is identified, an optimal response strategy is preferentially matched by utilizing a graph similarity, to support evolution and update of the early warning mechanism.

[0049] Compared with the prior art, the present application has the following beneficial effects:

[0050] The present application can perform high-order structure modeling on the packaging process state in a space-time-semantic three-dimensional mode by constructing a multi-dimensional tensor flow data structure and a packaging process semantic graph; further, by extracting a node semantic state vector, an endogenous disturbance response function is constructed to analyze a response ability of the system to a disturbance trend from a disturbance direction; potential instability signs can be identified in advance before obvious faults or production deviations occur, to realize a truly feed-forward early warning and prevention and control mechanism. A disturbance response weight dynamic adjustment mechanism is introduced to treat high-frequency light disturbances (such as mechanical light vibration, environmental light variation and action tolerance) in the system running differently;

[0051] By introducing the instant disturbance activation term and the historical disturbance accumulation term, the response intensity and false alarm suppression are automatically balanced; in actual production, the risk of triggering false alarms due to small errors can be greatly reduced, and the system stability and operator trust can be improved. The disturbance activation suppression reconciliation formula adjusts the system sensitivity in real time through the weight parameter in the intrinsic disturbance response function; the response sensitivity is enhanced when the real potential risk comes; during continuous light disturbance, the system can automatically converge to reduce the response intensity; the system has the unified ability of emergency and stability, and provides dynamic adaptability for complex production lines. The semantic elements in the image are extracted by using a deep neural network, such as workstation type, operation object, action state, spatial position, time information, etc.; and an interpretable semantic atlas structure is constructed through a semantic state vector; on the basis of structured expression, traceable, interpretable and reconfigurable early warning decision support is realized;

[0052] Through the region triggering mechanism based on image difference, the pixel change proportion of each region is calculated in real time, the process change region is accurately located, the sampling frequency is dynamically enhanced, low-redundancy high-gain sampling is realized, and the resource utilization efficiency of the monitoring system is significantly improved. The time decay semantic guidance factor dynamic adjustment mechanism based on trigger memory is introduced, the historical sampling data and smooth time decay are combined, the sampling weight of each region is effectively adjusted, the hot region is continuously concerned, the long-term static region is weakened, and the intelligent level of the sampling strategy is improved. Through the dynamic sampling and semantic guidance mechanism, small and sudden changes in the state of the packaging production line are effectively captured, the detection sensitivity and response speed of the abnormal state are improved, and the production safety and product quality are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A dynamic visual monitoring packaging production line early warning method process schematic diagram of the application;

[0054] Figure 2 A dynamic visual monitoring packaging production line early warning system structure schematic diagram of the application;

[0055] Figure 3 A packaging process semantic atlas acquisition method process schematic diagram provided by the application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0057] Embodiment 1

[0058] Please refer toFigure 1 and Figure 3 As shown, this embodiment further illustrates the early warning method for a packaging production line under dynamic visual monitoring proposed in this invention, including:

[0059] With the widespread application of automated packaging production lines, real-time monitoring and anomaly early warning of the production process have become crucial means to ensure product quality and production efficiency. Traditional early warning methods for packaging production lines mainly rely on visual monitoring systems, using technologies such as image recognition and motion detection to monitor the packaging process status. However, existing technologies still have many shortcomings and are difficult to meet the needs of complex actual production environments.

[0060] Existing early warning methods typically employ fixed parameter configurations, failing to dynamically adjust their sensitivity to changes in packaging process conditions based on the actual on-site scenario. This leads to frequent system responses under normal fluctuations or insufficient responses to potential instability trends, severely impacting the stability and reliability of the early warning system. Packaging production sites experience various non-abnormal natural disturbances, such as mechanical micro-vibrations and changes in ambient lighting. While these disturbances may cause visual differences, they do not necessarily indicate a risk in the production process. Traditional methods struggle to effectively distinguish between such normal disturbances and potential abnormal evolution, generally employing uniform threshold processing, resulting in a high false alarm rate and reducing the system's reliability and practical value.

[0061] Most existing systems focus on analyzing the state within a single frame or short time window, lacking the ability to learn from and suppress historical disturbance trends. Faced with continuous and high-frequency disturbance signals, these systems are prone to over-response, lacking self-regulation and suppression mechanisms, thus reducing the robustness and stability of early warnings. Furthermore, most early warning methods treat changes in packaging state as isolated events, ignoring the temporal correlation and causal trajectory of state evolution. They fail to capture the gradually accumulating instability trends and linked anomalies, such as the frame-by-frame tilting of the packaging box until it tipes over, or gradual anomalies like deceleration and positional drift caused by equipment aging.

[0062] Actual production environments are dynamic and constantly changing, and packaging processes involve factors such as task switching, production cycle adjustments, and manual intervention. Existing early warning systems primarily rely on real-time status assessments, lacking intelligent backtracking and strategy optimization capabilities that combine historical intervention effects and scenario evolution records. This prevents the system from automatically optimizing its response strategy in recurring abnormal scenarios, limiting both the efficiency and accuracy of early warning processing.

[0063] In view of this, the present invention proposes an early warning method for a packaging production line under dynamic visual monitoring, comprising:

[0064] S1. Continuously acquire images of the packaging production line to obtain image sequences of the packaging production line; extract semantic feature information representing the packaging process status from the image sequences, and construct a multidimensional tensor flow data structure based on the semantic feature information to generate a packaging process semantic map;

[0065] S2. Introduce semantic behavior curvature theory to perform curvature analysis on the semantic graph of packaging process, calculate the rate of change of semantic evolution curvature in different time segments, and use the sliding window dynamic curvature detection mechanism to identify abnormal curvature regions and generate semantic curvature anomaly information.

[0066] S3. Construct an internal disturbance response function based on the current running state vector of the node in the packaging process semantic graph, calculate the response sensitivity of the internal disturbance response function in the preset disturbance direction, and output the internal disturbance energy index representing the potential instability trend of the current packaging state.

[0067] S4. Perform synchronous fusion analysis on semantic bending anomaly information and internal driving disturbance energy index, determine the risk level of packaging production line based on fusion results, and generate corresponding graded early warning signals.

[0068] S5. Trigger corresponding response strategies based on graded early warning signals, and generate semantic perturbation memory maps based on the response strategies; when identifying semantic bending behaviors or energy perturbation trends with similar structures, prioritize matching the optimal response strategy based on map similarity to support the evolution and update of the early warning mechanism.

[0069] Methods for obtaining image sequences include:

[0070] Industrial vision cameras are deployed at different key workstations on the packaging production line to continuously capture images of the packaging operation process. The industrial vision cameras are set up through fixed installation or robotic arm tracking, and a capture timestamp is added to each frame of the image.

[0071] Key workstations include feeding station, filling station, calibration station, packaging station, labeling station, coding station, inspection station, boxing station, and sorting station; image sequence acquisition is performed in continuous frame capture mode, with the acquisition frequency set according to the packaging production line cycle time, and a ring buffer structure is used to buffer video image data;

[0072] It should be noted that the feeding station's function is to supply materials to be packaged, containers, labels, packaging materials, etc., to the main line; key visual features include material arrival status recognition and automatic positioning. The filling station (or filling / feeding station)'s function is to fill materials into packaging containers, such as boxes, bottles, bags, etc.; key visual features include filling amount, whether there is misalignment, and whether there is any omission.

[0073] Shaping / Correcting Station Function: Adjusts packages with irregular shapes or misaligned positions after filling; Visual Focus: Position calibration, orientation judgment. Packaging Station (Heat sealing, pressure sealing, folding sealing, etc.) Function: Performs preliminary sealing treatment on the packaging; Visual Focus: Seal integrity, edge misalignment.

[0074] Labeling station function: Automatically applies labels or barcodes; visual focus: label position, skewness, missing labels, or re-labeling. Coding / inkjet printing station function: Prints production batch, date, QR code, and other information; visual focus: character clarity, positional offset.

[0075] Checkweighing / Inspection Station Function: Used for online detection of parameters such as weight, appearance, and dimensions; Visual Focus: Visual recognition combined with weighing or 3D contour detection. Sealing / Packaging Station Function: Seals individual packages and packs them into outer boxes or cartons; Visual Focus: Packaging, stacking status, and whether the box seal is closed. Rejection / Sorting Station Function: Rejects or diverts defective products, empty packages, overweight products, and products with abnormal labeling; Visual Focus: Accurate linkage between matching warning signals and visual recognition results.

[0076] A region-triggered mechanism based on image difference is introduced, dividing the image into m regions, and calculating the pixel change ratio of the image difference map for each region; the pixel change ratio is... ;in, Indicates the first Semantic-guided differential response values ​​for the region; Indicates the region in the current frame No. The brightness value of each pixel; Indicates the region in the previous frame No. The brightness value of each pixel; Indicates the total number of pixels within the region; Indicates the first The semantic guiding factor corresponding to the region; Indicating an index for a region; based on expert experience, The value ranges from 0 to 1;

[0077] It should be noted that the operating status and environmental conditions of packaging production lines are often in a state of dynamic change, such as adjustments to process parameters, differences in equipment operation, or sudden abnormal events. These changes cause the semantic importance of different areas to change continuously over time. Static, fixed semantic guidance factors cannot accurately reflect actual monitoring needs, easily leading to resource waste or the risk of missed detections. Sampling and computing resources are limited. If all areas are assigned the same or fixed weights, it will inevitably lead to the collection and processing of a large amount of invalid data, reducing the overall efficiency of the system.

[0078] Therefore, by triggering the memory adjustment mechanism, the first The semantic guidance factor corresponding to the region is dynamically adjusted. ;in, Indicates the area At the point of time Semantic guiding factor; Indicates the area At the point of time Semantic guiding factor; This represents the time decay coefficient, which controls the gradual weakening of historical influence over time. This represents the trigger enhancement coefficient, which provides an additional stimulus to the factor when the region was sampled at the previous time point; Indicates the region at the previous time point Whether it has been sampled: 1 if sampled, 0 otherwise.

[0079] It should be noted that over time, if a region remains unsampled (triggered) for an extended period, the semantic guidance factor will gradually decay, and the sampling priority will decrease. This prevents resources from being concentrated in inactive regions for extended periods, improving the utilization rate of sampling resources. When a region is sampled ( When a factor is added with a fixed incentive to reflect recent activity in that area, its sampling weight is appropriately increased to ensure continuous attention to dynamic hotspot areas. By combining historical factors and real-time trigger information, the system can smoothly adapt to dynamic changes in areas, avoiding sudden changes in sampling strategies and improving system stability. This formula design combines a time-decreasing forgetting mechanism with memory enhancement of trigger events, ensuring both the flexibility and stability of dynamic allocation of sampling resources and focusing on real-time changes in hotspot areas. It has strong practicality and innovative value, fully meeting the intelligent adjustment requirements of dynamic visual monitoring systems for sampling strategies.

[0080] When the pixel change ratio of any region exceeds the preset pixel change ratio threshold, it is determined that the region is in a process change state, triggering dynamic enhancement sampling of the image frame with the process change; and performing noise filtering, brightness normalization and frame alignment operations on the image sequence to generate a standardized image sequence.

[0081] This solution addresses the following problems with existing technologies: Traditional visual monitoring systems for packaging production lines often employ a fixed-frequency, uniform acquisition method across the entire area, failing to dynamically adjust sampling strategies based on production status. This results in a large amount of invalid image data and significant resource waste. Existing technologies struggle to promptly detect dynamic changes in different critical workstations and areas within the packaging process, hindering the rapid identification of process anomalies or sudden changes in status, thus affecting the timeliness and accuracy of anomaly warnings.

[0082] Traditional methods typically assign static or empirical weights to monitoring areas, lacking the ability to adaptively adjust the semantic importance of areas to dynamically change with the production environment. This results in insufficient sampling frequency in key change areas and unsatisfactory anomaly detection. The system collects a large number of redundant images, increasing storage and computational burden and reducing the real-time performance and response speed of the overall monitoring system. Existing systems fail to adequately filter noise, normalize brightness, and align frames in image sequences, affecting the accuracy of subsequent semantic feature extraction and consequently impacting the accurate identification of packaging process conditions and the detection of abnormal bending.

[0083] Compared to existing technologies, the advantages are as follows: By using a region-triggered mechanism based on image difference, the pixel change ratio of each region is calculated in real time, accurately locating areas of process change, dynamically increasing the sampling frequency, achieving low-redundancy, high-gain sampling, and significantly improving the resource utilization efficiency of the monitoring system. A dynamic adjustment mechanism based on trigger memory and time decay semantic guidance factors is introduced, combining historical sampling data and smooth time decay to effectively adjust the sampling weight of each region, ensuring continuous attention to hot spots, weakening long-term static areas, and improving the intelligence level of the sampling strategy. Through dynamic sampling and semantic guidance mechanisms, minute and sudden changes in the state of the packaging production line are effectively captured, improving the detection sensitivity and response speed of abnormal states, and ensuring production safety and product quality.

[0084] Methods for obtaining the semantic graph of packaging processes include:

[0085] A semantic recognition model based on deep neural networks is used to detect and classify semantic elements of packaging processes in image sequences. The semantic recognition model can be either a deep convolutional neural network or an object detection network model. The semantic elements include workstation type, operation object, action state, position coordinates, and time information.

[0086] The recognition results are encoded into fixed-length semantic state vectors through one-hot encoding, numerical normalization, and positional encoding. The vectors are then organized according to the temporal order of the image frames and the spatial order of the workstations to construct a multi-dimensional tensor flow data structure that includes temporal, spatial, and semantic dimensions.

[0087] Each fixed-length semantic state vector in the multidimensional tensor flow data structure is mapped to a node in the semantic graph. Edge connections between nodes are established based on temporal evolution relationships and operational causal relationships, forming a packaging process semantic graph that reflects the semantic structure and evolution path of the packaging production process.

[0088] It should be noted that the temporal evolution relationship refers to the continuous state transition relationship between the same workstation or the same object at different points in time, reflecting the natural evolution of behavior over time during the packaging process. For example, the same packaging bottle is in the state of "waiting to be filled" at 10:00:00, "filling" at 10:00:01, and then "filling completed" at 10:00:02; at the same workstation G1, the previous frame was "no bottle", and the current frame shows a bottle entering, which belongs to the state evolution.

[0089] The causal relationship in operational processes refers to the fact that the completion of one process action is a prerequisite or necessary condition for another process action, reflecting the logical dependencies across workstations or objects in the packaging production line. For example, "filling completed" at workstation G1 → "sealing started" at workstation G2; "labeling completed" → "visual quality inspection" → "packing and warehousing"; if the status of a certain operation object is "missing", it will cause the downstream process to skip.

[0090] Methods for obtaining the rate of change of semantic evolution curvature in different time segments include:

[0091] In the semantic graph of packaging process, a sequence of continuous semantic state nodes under the same object or the same workstation is selected along the edge of the time evolution relationship; based on the sequence of continuous semantic state nodes, the semantic state vector trajectory is extracted and mapped to the semantic state space to form a semantic evolution path;

[0092] The three-point discrete curvature estimation is used to calculate the curvature of the semantic evolution path. The semantic state vectors corresponding to three consecutive semantic state nodes are differentially analyzed to calculate the curvature of the intermediate node relative to the nodes before and after it. Within a preset time window, the rate of change of the curvature over time is statistically analyzed to obtain the rate of change of semantic evolution curvature in different time segments.

[0093] Methods for obtaining semantic curvature anomaly information include:

[0094] The semantic evolution curvature change rate of each time segment is arranged in chronological order to form a semantic evolution curvature change rate sequence; a fixed-length sliding time window is preset and slides on the semantic evolution curvature change rate sequence with a preset step size to extract the curvature change rate segment corresponding to each sliding window;

[0095] For each sliding window segment, calculate the statistical feature index of the rate of change of semantic evolution curvature within the sliding window. The statistical feature index includes the average rate of change of curvature, the maximum rate of change of curvature, and the standard deviation of the rate of change of curvature. Then compare the statistical feature index with the preset threshold for judging abnormal statistical feature index.

[0096] When the statistical feature index exceeds the preset statistical feature index threshold, it is determined that there is an abnormal bending trend in the semantic state evolution of the time period corresponding to the current window; adjacent or overlapping abnormal windows are merged to form an abnormal interval, and the start and end times, the set of associated semantic state nodes and the corresponding abnormal intensity index of the abnormal interval are output to constitute semantic bending abnormal information.

[0097] Methods for constructing the response function of an internally driven disturbance include:

[0098] Select a preset target node from the packaging process semantic graph. The preset target node corresponds to the current operating status of any key workstation in the packaging production line. Extract the current operating status vector of the preset target node. The operating status vector includes workstation type code, operation object information, action parameters, spatial location information and time label.

[0099] A set of perturbation direction vectors is predefined to characterize potential changes in packaging processes. These perturbation direction vectors include process parameter perturbation vectors, position perturbation vectors, and semantic category perturbation vectors.

[0100] For each disturbance direction, based on the linear projection relationship between the current node's running state vector and the disturbance direction and the state disturbance gradient, the response sensitivity of the running state vector under the disturbance direction is calculated, and the intrinsic disturbance response function is constructed.

[0101] It should be noted that linear projection refers to the projection of the current state vector onto a perturbation direction vector, used to measure the tendency of the state to change in that perturbation direction. The state perturbation gradient represents the rate of change of the overall system response function (such as the stability evaluation function) after the current state vector is perturbed in the perturbation direction.

[0102] The internal disturbance response function is: ;in, Indicates the direction vector of the disturbance The response intensity (i.e., the value of the internal driving disturbance response function); Indicates the first One perturbation direction vector; Indicates a point in time The semantic state vector; This represents the vector dot product operation; Represents the vector along the direction of the disturbance. The state gradient; Denotes the squared L2 norm of a vector; Indicates the disturbance response weighting coefficient;

[0103] It should be noted that in the early warning process of a packaging production line under dynamic visual monitoring, the intrinsic disturbance response function is used to measure the sensitivity of the current packaging process state to potential disturbances, and its disturbance response weight coefficient is a key parameter controlling the strength of the response. However, packaging production lines experience various normal disturbances and fluctuations, such as slight mechanical vibrations and changes in ambient light. If these brief and frequent disturbances are indiscriminately given high weights, it can easily lead to frequent false alarms, affecting system stability and operator confidence. The production line environment is complex and variable, and disturbance signals contain noise and uncertainty; a single fixed weight cannot meet the needs of all operating conditions.

[0104] Based on this, the disturbance response weight coefficients in the internal driving disturbance response function are adaptively adjusted.

[0105] The perturbation response weight coefficients are adaptively adjusted using the perturbation activation suppression harmonic formula, which is as follows: ;in, Indicates the adaptive adjustment at time point The disturbance response weighting coefficient; Indicates at a point in time The current perturbation gradient strength of a node in the packaging process semantic graph is specifically the gradient norm of the intrinsic perturbation response function. ; The historical disturbance backtracking window size represents the range of disturbance response coefficients considered over the historical time dimension; Represents a time step index variable; Indicates the current time point Forward The perturbation gradient strength at each time unit (e.g., frame, second, sampling point); This represents the instantaneous disturbance activation adjustment factor, which controls the amplification effect of the current disturbance intensity on the disturbance response weighting coefficient. This represents the cumulative suppression factor for historical disturbances, indicating the degree to which the system suppresses the total amount of historical disturbances. An index representing a specific point in time; based on expert experience. and The value ranges from 0 to 1;

[0106] This solution addresses the following problems with existing technologies: Most existing methods use fixed parameter configurations, failing to dynamically adjust sensitivity to changes in packaging process conditions according to the scenario. This leads to the system overreacting to normal fluctuations or being insensitive to potential instability warnings, affecting system stability and reliability. Existing systems rely on single visual features such as image recognition or motion detection, triggering warnings as soon as a change is detected. However, in actual packaging processes, many disturbances fall within the scope of normal fluctuations, such as various non-abnormal natural disturbances in the packaging production area, including mechanical micro-vibrations and lighting fluctuations. While these disturbances can cause visual differences, they do not necessarily indicate system risk. Traditional methods cannot effectively distinguish between normal disturbances and potential abnormal evolution, often treating all changes uniformly, resulting in high false alarm rates and decreased system reliability. These disturbances are characterized by being short-lived, frequent, and low-risk, but traditional methods struggle to effectively identify their differences, easily leading to frequent false alarms and operational interference.

[0107] Existing mechanisms often focus on the state of the current frame, lacking trend learning or suppression of historical disturbances. This leads to over-response when faced with continuous high-frequency disturbances, a lack of self-inhibition capabilities, and reduced robustness of the early warning system. Many early warning systems treat packaging state changes as isolated events, relying on the state of a single frame or a short period to determine the presence of anomalies, lacking modeling and utilization of state evolution trends and causal trajectories. This results in the system's inability to identify gradually accumulating instability trends, such as packaging boxes tilting frame by frame until they tip over; it also fails to capture the interconnected faults gradually triggered by multiple slight deviations, ignoring progressive abnormal behaviors during production such as position drift and deceleration. In actual production, packaging processes involve dynamic changes such as task switching, cycle time variations, equipment aging, and manual intervention. In existing technologies, most early warning strategies make judgments based on the current state in real time, rather than combining historical intervention effects and scenario evolution records for intelligent backtracking and strategy optimization. This results in response strategies failing to automatically optimize or evolve, requiring recalculation in repetitive scenarios, making it difficult to improve processing efficiency and accuracy.

[0108] Compared to existing technologies, the advantages are as follows: By constructing a multidimensional tensor flow data structure and a packaging process semantic graph, a high-order structural model of the packaging process state is performed in the three dimensions of space, time, and semantics; furthermore, by extracting node semantic state vectors, an intrinsic disturbance response function is constructed to analyze the system's response capability to disturbance trends from the perspective of disturbance direction; potential instability signs can be identified in advance before obvious faults or production deviations occur, realizing a truly feedforward early warning and prevention mechanism. A dynamic adjustment mechanism for disturbance response weights is introduced to differentiate between high-frequency minor disturbances in system operation (such as minor mechanical vibrations, ambient light changes, and action tolerances);

[0109] By introducing immediate disturbance activation terms and historical disturbance accumulation terms, the system automatically balances response strength and false alarm suppression. In actual production, this significantly reduces the risk of false alarms triggered by small errors, improving system stability and operator confidence. The disturbance activation suppression harmonic formula adjusts system sensitivity in real time through weight parameters in the intrinsic disturbance response function, enhancing response sensitivity when real potential risks arise. During periods of sustained minor disturbances, the system automatically converges, reducing response strength, ensuring a unified capability of emergency preparedness and stability, and providing dynamic adaptability for complex production lines. Deep neural networks are used to extract semantic elements from images, such as workstation type, operating object, action state, spatial location, and temporal information. An interpretable semantic graph structure is constructed using semantic state vectors, enabling traceable, interpretable, and reconfigurable early warning decision support based on structured representation.

[0110] Methods for obtaining the internal driving disturbance energy index include:

[0111] Based on the constructed intrinsic disturbance response function, the response sensitivity of the preset target node under different disturbance direction vectors is calculated. The intrinsic disturbance response function is used to quantify the degree of response of the current running state vector to semantic state changes in each disturbance direction.

[0112] The quantization process includes multiplying the linear projection of the current operating state vector onto the disturbance direction vector with the state disturbance gradient in that disturbance direction to obtain the response intensity value in each disturbance direction; normalizing the response intensity values ​​in each disturbance direction to obtain the overall disturbance response sensitivity index of the current packaging state, which serves as the internal disturbance energy index representing the potential instability trend of the current packaging state.

[0113] Methods for determining the risk level of a packaging production line and generating corresponding graded early warning signals include:

[0114] Semantic curvature anomaly information and intrinsic disturbance energy index are synchronized and time-series aligned, matched and normalized within a unified time window; a fusion function is constructed to weight and fuse the normalized semantic curvature anomaly information and intrinsic disturbance energy index according to preset weights to generate a comprehensive risk index.

[0115] Based on comprehensive risk indicators and preset risk thresholds, the current status of the packaging production line is divided into different risk levels, including normal, warning, severe, and critical levels. According to the determined risk level, a corresponding graded early warning signal is generated, which includes risk level information, abnormal time range, abnormal type, and identification of relevant workstations.

[0116] Methods that utilize graph similarity to prioritize matching the optimal response strategy include:

[0117] Based on the risk level corresponding to the graded early warning signal, a matching response strategy is triggered. The response strategy includes production rhythm adjustment, key workstation suspension, manual intervention reminder, and execution of process correction instructions. During the execution of the response strategy, response events within the current time period are recorded. The response events include the packaging process semantic graph structure subgraph within the current time window, the corresponding semantic bending anomaly information, the internal driving disturbance energy index, the risk level, and the triggered matching response strategy.

[0118] The response events are structured and organized into a semantic perturbation memory graph, which stores the correlation between historical warnings and response events. Substructures in the current semantic perturbation memory graph are identified, and based on the subgraph isomorphism detection algorithm, the substructure is compared with the subgraphs already recorded in the semantic perturbation memory graph and the similarity is calculated.

[0119] When the similarity between the current semantic bending behavior and the semantic bending behavior of a subgraph recorded in the historical semantic perturbation memory map is greater than a preset similarity threshold, the optimal response strategy corresponding to the historical semantic perturbation memory map is immediately invoked first to perform early warning response strategy invocation and dynamic optimization update.

[0120] The preset pixel change ratio threshold is set by staff based on historical data analysis results. This historical analysis process includes the system collecting multiple pixel change ratios and calculating their average value as a reference to obtain the preset pixel change ratio threshold. Similarly, preset abnormal statistical feature indicator judgment thresholds, preset statistical feature indicator thresholds, preset risk thresholds, and preset similarity thresholds are also set by staff based on the system's historical operating data and specific application scenario requirements. These preset thresholds can be adjusted by staff during system operation according to actual conditions.

[0121] This embodiment constructs a multidimensional tensor flow data structure and a packaging process semantic graph to perform high-order structural modeling of the packaging process state in three dimensions: space, time, and semantics. Furthermore, by extracting node semantic state vectors, it constructs an intrinsic disturbance response function to analyze the system's response capability to disturbance trends from the perspective of disturbance direction. This allows for the early identification of potential instability signs before obvious faults or production deviations occur, achieving a truly feedforward-type early warning and prevention mechanism. A dynamic adjustment mechanism for disturbance response weights is introduced to differentiate between high-frequency minor disturbances in system operation (such as minor mechanical vibrations, ambient light changes, and action tolerances).

[0122] By introducing immediate disturbance activation terms and historical disturbance accumulation terms, the system automatically balances response strength and false alarm suppression. In actual production, this significantly reduces the risk of false alarms triggered by small errors, improving system stability and operator confidence. The disturbance activation suppression harmonic formula adjusts system sensitivity in real time through weight parameters in the intrinsic disturbance response function, enhancing response sensitivity when real potential risks arise. During periods of continuous minor disturbances, the system automatically converges, reducing response strength, ensuring a unified capability of emergency preparedness and stability, and providing dynamic adaptability for complex production lines. Deep neural networks are used to extract semantic elements from images, such as workstation type, operating object, action state, spatial location, and temporal information. An interpretable semantic graph structure is constructed using semantic state vectors. Based on this structured expression, traceable, interpretable, and reconfigurable early warning decision support is achieved.

[0123] By employing a region-triggered mechanism based on image difference, the system calculates the pixel change ratio of each region in real time, accurately locates areas of process change, dynamically enhances the sampling frequency, and achieves low-redundancy, high-gain sampling, significantly improving the resource utilization efficiency of the monitoring system. A dynamic adjustment mechanism based on trigger memory and time decay semantic guidance factors is introduced. Combining historical sampling data and smooth time decay, this effectively adjusts the sampling weights of each region, ensuring continuous attention to hotspot areas and weakening long-term static areas, thus improving the intelligence level of the sampling strategy. Through dynamic sampling and semantic guidance mechanisms, the system effectively captures minute and sudden changes in the packaging production line status, improving the detection sensitivity and response speed of abnormal states, and ensuring production safety and product quality.

[0124] Example 2

[0125] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A dynamic visual monitoring-based early warning system for a packaging production line is provided, comprising:

[0126] The multidimensional visual semantic module continuously acquires images of the packaging production line to obtain image sequences of the packaging production line; extracts semantic feature information representing the packaging process status from the image sequences, and constructs a multidimensional tensor flow data structure based on the semantic feature information to generate a packaging process semantic map;

[0127] The semantic curvature analysis module introduces the semantic behavior curvature theory to perform curvature analysis on the semantic map of packaging process, calculates the rate of change of semantic evolution curvature in different time segments, and uses a sliding window dynamic curvature detection mechanism to identify abnormal curvature regions and generate semantic curvature anomaly information.

[0128] The internal disturbance energy calculation module constructs an internal disturbance response function based on the current operating state vector of the node in the packaging process semantic graph, calculates the response sensitivity of the internal disturbance response function in the preset disturbance direction, and outputs an internal disturbance energy index that represents the potential instability trend of the current packaging state.

[0129] The collaborative risk assessment module performs synchronous fusion analysis on semantic bending anomaly information and internal driving disturbance energy index, determines the risk level of the packaging production line based on the fusion results, and generates corresponding graded early warning signals.

[0130] The early warning response evolution module triggers corresponding response strategies based on the graded early warning signals and generates a semantic perturbation memory map based on the response strategies. When it identifies semantic bending behavior or energy perturbation trends with similar structures, it uses the similarity of the map to prioritize matching the optimal response strategy, supporting the evolution and update of the early warning mechanism.

[0131] Since the electronic device described in this embodiment is the one used in implementing the packaging production line early warning method and system under dynamic visual monitoring in the embodiments of this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the packaging production line early warning method and system under dynamic visual monitoring described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. As long as those skilled in the art implement the electronic device used in the packaging production line early warning method and system under dynamic visual monitoring in the embodiments of this application, it falls within the scope of protection of this application.

[0132] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0133] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for early warning of a packaging production line under dynamic visual monitoring, characterized in that, include: S1. Continuously acquire images of the packaging production line to obtain an image sequence of the packaging production line; Semantic features representing the packaging process status are extracted from image sequences, and a multidimensional tensor flow data structure is constructed based on the semantic features to generate a packaging process semantic graph. S2. Introduce semantic behavior curvature theory to perform curvature analysis on the semantic graph of packaging process, calculate the rate of change of semantic evolution curvature in different time segments, and use the sliding window dynamic curvature detection mechanism to identify abnormal curvature regions and generate semantic curvature anomaly information. S3. Construct an internal disturbance response function based on the current running state vector of the node in the packaging process semantic graph, calculate the response sensitivity of the internal disturbance response function in the preset disturbance direction, and output the internal disturbance energy index representing the potential instability trend of the current packaging state. S4. Perform synchronous fusion analysis on semantic bending anomaly information and internal driving disturbance energy index, determine the risk level of packaging production line based on fusion results, and generate corresponding graded early warning signals. S5. Trigger corresponding response strategies based on graded early warning signals, and generate semantic perturbation memory maps based on the response strategies; when identifying semantic bending behaviors or energy perturbation trends with similar structures, prioritize matching the optimal response strategy based on map similarity to support the evolution and update of the early warning mechanism.

2. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 1, characterized in that, The method for obtaining the image sequence includes: Industrial vision cameras are deployed at different key workstations on the packaging production line to continuously acquire images of the packaging operation process. A region triggering mechanism based on image difference is introduced, which divides the image into m regions. The pixel change ratio of the image difference map is calculated for each region. When the pixel change ratio of any region exceeds the preset pixel change ratio threshold, it is determined that the region is in a process change state, and dynamic enhancement sampling of the image frame with the process change is triggered. Noise filtering, brightness normalization and frame alignment operations are performed on the image sequence to generate a standardized image sequence.

3. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 2, characterized in that, The method for obtaining the semantic graph of the packaging process includes: A semantic recognition model based on deep neural networks is used to detect and classify the semantic elements of packaging processes involved in image sequences, encode them into fixed-length semantic state vectors, map each fixed-length semantic state vector to a node in the semantic graph, and establish edge connections between nodes based on temporal evolution and causal relationships of operational processes to form a semantic graph of packaging processes.

4. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 3, characterized in that, The methods for obtaining the semantic evolution curvature change rate for different time segments include: In the semantic graph of packaging process, a sequence of continuous semantic state nodes under the same object or the same workstation is selected along the temporal evolution relationship edge; the semantic state vector trajectory is extracted to form a semantic evolution path; the curvature of the semantic evolution path is calculated to obtain the semantic evolution curvature change rate of different time segments.

5. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 4, characterized in that, The method for obtaining the semantic curvature anomaly information includes: The semantic evolution curvature change rate of each time segment is arranged in chronological order to form a semantic evolution curvature change rate sequence; statistical feature indicators are calculated and compared with preset abnormal statistical feature indicator judgment thresholds; when the preset statistical feature indicator thresholds are exceeded, it is determined that there is an abnormal curvature trend in the semantic state evolution of the time segment corresponding to the current window, and semantic curvature abnormal information is constituted.

6. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 5, characterized in that, The method for constructing the intrinsic disturbance response function includes: Select a preset target node from the semantic graph of packaging process, extract the current operating state vector of the preset target node, and preset a set of disturbance direction vectors to characterize potential changes in packaging process; for each disturbance direction, calculate the response sensitivity of the operating state vector under the disturbance direction, construct the internal disturbance response function, and adaptively adjust the disturbance response weight coefficient in the internal disturbance response function.

7. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 6, characterized in that, The method for obtaining the internal driving disturbance energy index includes: Based on the constructed intrinsic disturbance response function, the response sensitivity of the preset target node under different disturbance direction vectors is calculated. The response degree of the current operating state vector to semantic state changes in each disturbance direction is quantified by the intrinsic disturbance response function and normalized to obtain the overall disturbance response sensitivity index of the current packaging state, which serves as the intrinsic disturbance energy index representing the potential instability trend of the current packaging state.

8. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 7, characterized in that, The method for determining the risk level of the packaging production line and generating corresponding graded early warning signals includes: Semantic curvature anomaly information and internal driving disturbance energy index are synchronized and time-series aligned, matched and normalized within a unified time window; a fusion function is constructed to generate a comprehensive risk index; based on the comprehensive risk index and a preset risk threshold, the current state of the packaging production line is divided into different risk levels, and corresponding graded early warning signals are generated according to the determined risk level.

9. The method for early warning of a packaging production line under dynamic visual monitoring according to claim 8, characterized in that, The method of using graph similarity to prioritize matching the optimal response strategy includes: Based on the risk level corresponding to the graded early warning signal, a matching response strategy is triggered. During the execution of the response strategy, response events within the current time period are recorded. The response events are structured and organized to construct a semantic perturbation memory graph. Substructures in the current semantic perturbation memory graph are identified, and the substructures are compared with the subgraphs already recorded in the semantic perturbation memory graph for structural similarity calculation. When the similarity between the current semantic bending behavior and the semantic bending behavior of a subgraph recorded in the historical semantic perturbation memory map is greater than a preset similarity threshold, the optimal response strategy corresponding to the historical semantic perturbation memory map is immediately invoked first to perform early warning response strategy invocation and dynamic optimization update.

10. A packaging production line early warning system under dynamic visual monitoring, used to implement the packaging production line early warning method under dynamic visual monitoring as described in any one of claims 1 to 9, characterized in that, include: The multidimensional visual semantic module continuously acquires images of the packaging production line, obtains image sequences of the packaging production line, and generates a semantic map of the packaging process. The semantic curvature analysis module introduces semantic behavior curvature theory to perform curvature analysis on the semantic graph of packaging process and generate semantic curvature anomaly information. The internal disturbance energy calculation module constructs the internal disturbance response function and outputs the internal disturbance energy index, which represents the potential instability trend of the current packaging state. The collaborative risk assessment module performs synchronous fusion analysis of semantic bending anomaly information and internal driving disturbance energy index to determine the risk level of the packaging production line and generate graded early warning signals. The early warning response evolution module triggers corresponding response strategies based on the graded early warning signals, generating a semantic perturbation memory map; Prioritize matching the optimal response strategy and support the evolution and updating of the early warning mechanism.

Images