Intelligent monitoring method and system for quality of feed packaging bag based on sensor network

Abstract

The application discloses a kind of based on sensor network's feed packaging bag quality intelligent monitoring method and system, including the following steps: acquisition feed packaging process multi-source quality monitoring data and equipment operating state data, form packaging quality time series data;Data are time aligned and packaging process division, obtain packaging quality state sequence;According to sensor node attribute and packaging process change, construct packaging quality dynamic correlation graph;State sequence and dynamic correlation graph are input GDN graph deviation network, and node quality state prediction result is obtained by process constraint graph attention aggregation;According to prediction result and normal packaging correlation benchmark, determine state deviation and correlation deviation, form packaging bag quality deviation feature;Further identify abnormal node, quality abnormal type and abnormal action area, and generate packaging bag quality monitoring result.The application can improve packaging quality abnormal identification and positioning ability.

Description

Technical Field

[0001] This invention relates to the field of feed packaging quality monitoring technology, specifically to a method and system for intelligent monitoring of feed packaging bag quality based on sensor networks. Background Technology

[0002] Feed packaging is a crucial step in feed production, storage, and transportation. The forming state, filling state, sealing quality, and conveying stability of the packaging bags directly affect the sealing performance, storage safety, and subsequent distribution of feed products. Feed packaging typically involves multiple continuous processes, including bag unfolding, quantitative filling, sealing and pressing, and finished product conveying. These processes are strongly interconnected. For example, changes in bag tension can affect the bag's forming position and sealing stability; the coordination between heat-sealing temperature and pressure affects the tightness of the sealing area; and changes in filling weight can further affect the bag's stress and vibration during conveying. Therefore, the quality of the packaging bag is not determined independently by a single equipment parameter or a single sensor state, but rather by the coordinated changes of multiple sensor nodes throughout the various packaging processes.

[0003] Current methods for monitoring feed packaging quality typically involve manual sampling, single-point sensor threshold judgment, or post-packaging visual inspection. Manual sampling relies on operator experience, has a long inspection cycle, and struggles to provide real-time coverage of the continuous packaging process. Single-point threshold judgment primarily identifies anomalies based on whether a single parameter such as temperature, pressure, weight, or vibration exceeds a set range. While it can detect some obvious anomalies, it struggles to identify situations where a single sensor reading remains within the normal range while the correlation between multiple sensors has undergone abnormal changes. Post-packaging visual inspection usually only detects existing surface defects such as wrinkles, damage, and seal misalignment, lacking the ability to proactively identify process deviations that gradually develop during packaging.

[0004] With the development of sensor network technology, various sensors can be deployed at different operating locations on feed packaging equipment to continuously collect data on the forming, filling, sealing, and conveying processes. However, different sensors have different sampling characteristics and process response patterns, and the degree of influence of the same sensor on packaging quality varies significantly under different packaging processes. For example, in the filling stage, there is a strong correlation between weight changes, bag tension, and position changes, while in the sealing stage, the correlation between heat sealing temperature, heat sealing pressure, and the state of the sealing area is more important. If a fixed sensor correlation is always used for quality analysis, it is difficult to accurately reflect the changes in the node interaction relationships caused by the switching of packaging processes.

[0005] Existing methods based on machine learning or temporal anomaly detection can learn normal operation patterns from multivariate monitoring data, but they typically focus on the numerical changes of each variable or the relationships between fixed variables, failing to adequately consider the dynamic changes in sensor-related structures caused by changes in packaging processes. When a packaging quality anomaly manifests as a decrease in the correlation strength between multiple sensor nodes, changes in the associated objects, or an imbalance in the cooperative relationship, relying solely on a single deviation between the predicted and actual values ​​can easily lead to missed detections. Furthermore, existing methods typically interpret anomaly results only at the anomaly score level, making it difficult to further pinpoint the sensor nodes associated with the anomaly, the corresponding packaging processes, and the areas of abnormal effect on the packaging bag.

[0006] Therefore, how to provide a method and system for intelligent monitoring of feed packaging bag quality based on sensor networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent monitoring method and system for feed packaging bag quality based on sensor networks. This invention continuously collects multi-source quality monitoring data and packaging equipment operating status data during the feed packaging process, constructs a packaging quality status sequence corresponding to each packaging stage, and forms a dynamic correlation graph of packaging quality that changes with the packaging process based on sensor node attributes, node state change relationships, and the quality impact relationships of the packaging process. Furthermore, it utilizes a GDN graph deviation network to aggregate process constraint graph attention features of the dynamic correlation, obtaining sensor node quality status prediction results. Combining sensor state deviation and sensor correlation deviation, it forms packaging bag quality deviation features, thereby enabling the identification of abnormal sensor nodes, associated abnormal nodes, quality anomaly types, and abnormal action areas, and generating corresponding packaging bag quality monitoring results.

[0008] A method and system for intelligent monitoring of feed packaging bag quality based on sensor networks according to an embodiment of the present invention includes the following steps: S1 collects multi-source quality monitoring data and packaging equipment operation status data during the feed packaging process, and forms packaging quality time-series data according to packaging bag identification, node identification and collection time. S2, perform time alignment and process division on the packaging quality time series data, and encode the sensor data of each process stage to obtain the packaging quality state sequence; S3, acquire sensor node attributes and historical monitoring data of qualified packaging, establish the quality impact relationship of packaging process, construct a dynamic correlation diagram of packaging quality and continuously update it to obtain correlation change records; S4. Input the packaging quality status sequence and packaging quality dynamic correlation diagram into the GDN graph deviation network, and obtain the quality status prediction result through process constraint graph attention aggregation; S5. Establish a normal packaging correlation benchmark based on the historical monitoring data of qualified packaging, and calculate the dual deviation by combining the quality status prediction results and correlation change records to obtain the packaging bag quality deviation characteristics. S6. Identify abnormal sensor nodes and associated abnormal nodes based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, and determine the type of quality abnormality and the area of ​​abnormality. S7. Determine the quality status of the packaging bag based on the characteristics of the packaging bag quality deviation, and generate the packaging bag quality monitoring results by combining the identified quality anomaly types and the areas affected by the anomalies.

[0009] Optionally, S2 includes: S21, establish a unified time series based on the collection time in the packaging quality time series data, and match the time data of different sensors for the same packaging bag; S22, determine the stages of forming, loading, sealing and conveying processes based on the operating status of packaging equipment and changes in sensor response, and establish the correspondence between process stages and a unified time series; S23, extract sensor data according to the time range corresponding to each process stage to obtain process stage monitoring data under different process stages; S24. The process stage identifier and process stage monitoring data are jointly coded and arranged according to the sequence of processes to obtain the packaging quality status sequence.

[0010] Optionally, S3 includes: S31. Obtain the sensor type and installation location from the node configuration record, combine them with the packaging process stage to form node attribute characteristics, and extract the node temporal change characteristics from the packaging quality status sequence. S32, the node attribute features and node temporal change features are fused to obtain sensor node features, and the correlation degree of node state change is determined based on the node state change; S33, establish the quality impact relationship of the packaging process based on sensor node attributes and process stages, and select node combinations based on the correlation of node state changes to form an initial quality association set; S34. Determine the reliability of the quality association of node combinations based on historical monitoring data of qualified packaging, adjust the association weights in combination with process stages, construct a dynamic association diagram of packaging quality, and generate association change records.

[0011] Optionally, S34 includes: S341, Time alignment and process division are performed on the historical monitoring data of qualified packaging, and the state change relationship of node combinations in the initial quality association set is extracted to form historical node association data; S342, calculate the degree of common change, the degree of consistency of change direction, and the degree of association persistence of node combinations based on historical node association data to obtain historical association stability characteristics; S343, jointly encode the correlation degree of node state changes, the impact of packaging process quality, and historical correlation stability characteristics to obtain reliable quality correlation characteristics; S344: Based on the reliable characteristics of quality association and the current process stage, select reliable quality association relationships, adjust the node association weights, continuously construct a dynamic association diagram of packaging quality, and record association changes.

[0012] Optionally, S344 includes: S3441, Determine the effective sensor nodes based on the current process stage, and extract the node combination containing the effective sensor nodes from the initial quality association set; S3442, based on the reliable characteristics of quality association and the quality impact relationship of packaging process, filter out node combinations that do not match the current process to obtain reliable quality association relationships; S3443, determine the association weight of credible quality association based on the correlation degree of node state changes, historical association stability characteristics and the impact of packaging process quality; S3444 uses effective sensor nodes as graph nodes, reliable quality association relationships as graph edges, and association weights as graph edge attributes to continuously construct a dynamic association graph of packaging quality, and records association changes according to the monitoring time to obtain association change records.

[0013] Optionally, S4 includes: S41, Time alignment and process coding are performed on the historical monitoring data of qualified packaging to form a historical packaging quality status sequence, and a corresponding dynamic correlation graph and training samples are constructed to train the GDN graph deviation network. S42, extract the node time sequence status before the target monitoring time from the packaging quality status sequence, and determine the effective adjacent nodes and association weights based on the packaging quality dynamic association graph; S43, input the temporal state, association weight and current process stage of the target node and its effective adjacent nodes into the attention layer of the process constraint graph to obtain the association quality aggregation feature; S44 integrates the target node's temporal status, associated quality aggregation features, and the current process stage, and obtains the quality status prediction result for the target monitoring time through the trained GDN graph deviation network.

[0014] Optionally, S43 includes: S431, Extract the set of effective adjacent nodes and node association weights corresponding to the target sensor node based on the dynamic correlation diagram of packaging quality; S432, perform feature matching on the temporal states of the target node and each effective neighboring node to obtain attention response features that characterize the degree of correlation between node quality states; S433 integrates the attention response features with the node association weights and the packaging process quality influence relationship established in S3, suppresses the adjacent responses that do not match the current process, and obtains the process constraint attention weights. S434: Based on the process constraint attention weight, the temporal state of each effective adjacent node is weighted and aggregated to obtain the association quality aggregation feature of the target node.

[0015] Optionally, S5 includes: S51, compare the quality status prediction results with the actual monitoring status corresponding to the target monitoring time, and form the sensor status deviation based on the difference in node status prediction. S52, performs time alignment and process segmentation of historical monitoring data for qualified packaging, and adopts dynamic correlation with current packaging quality. Figure 1 A consistent composition method is used to establish a historical relationship diagram, forming a normal packaging relationship benchmark; S53, compare the current node association status with the normal packaging association benchmark, and combine the continuous association changes in the association change record to obtain the sensor association deviation; S54, jointly encodes the degree, direction and duration of sensor state deviation and sensor correlation deviation, retains node identifier, correlation relationship of associated nodes and time information, and obtains the packaging bag quality deviation characteristics.

[0016] Optionally, S6 includes: S61, using the trained GDN graph deviation network to predict the verification samples in the historical monitoring data of qualified packaging, determining the normal deviation range of nodes based on the prediction error, and screening abnormal sensor nodes by combining state deviation, correlation deviation and continuous state. S62, read the correlation relationship of abnormal sensor nodes along the dynamic correlation diagram of packaging quality, and identify the peer node with continuous abnormal correlation and the node with synchronous quality deviation change as the associated abnormal node. S63, based on the sensor type, installation location and process stage of the abnormal node, establish the correspondence between the sensor installation location and the effective area of ​​the packaging bag, and determine the abnormal process and abnormal action location; S64. Determine the quality anomaly type based on the quality impact relationship of the packaging process and the combination of abnormal nodes, and obtain the anomaly area by combining the location of the anomaly; no anomaly type or anomaly area is generated when no anomaly is detected.

[0017] The intelligent monitoring system for feed packaging bag quality based on sensor networks according to an embodiment of the present invention includes the following modules: The data acquisition module is used to collect multi-source quality monitoring data and packaging equipment operating status data to form packaging quality time-series data; The state sequence construction module is used to perform time alignment and process division on packaging quality time-series data to obtain a packaging quality state sequence. The dynamic correlation graph construction module is used to obtain sensor node attributes and historical monitoring data of qualified packaging, establish the quality impact relationship of packaging process, and construct a dynamic correlation graph of packaging quality and correlation change records. The GDN quality prediction module is used to train the GDN graph deviation network based on historical monitoring data of qualified packaging, and obtain the quality status prediction result through process constraint graph attention aggregation. The quality deviation analysis module is used to establish a normal packaging correlation benchmark, determine the sensor status deviation based on the quality status prediction results, determine the sensor correlation deviation based on the current node correlation status and correlation change records, and fuse them to obtain the packaging bag quality deviation characteristics. The anomaly location module is used to determine the abnormal sensor nodes and associated abnormal nodes based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, and to determine the quality anomaly type and the area of ​​anomaly when an anomaly exists. The quality result generation module is used to determine the quality status of packaging bags based on the characteristics of packaging bag quality deviations, and generate packaging bag quality monitoring results by combining the identified quality anomaly types and the areas affected by the anomalies.

[0018] The beneficial effects of this invention are: This invention continuously monitors the multi-source quality status of feed packaging through a sensor network, and organizes the data relationships of each sensor node according to the packaging process stage, so that the originally scattered monitoring information such as temperature, pressure, weight, tension, deformation and conveying status can be correlated and analyzed in a unified packaging quality status space, thereby improving the ability of multi-source sensor data to characterize the overall quality status of the packaging bag.

[0019] This invention determines reliable quality correlations based on the correlation degree of state changes between sensor nodes, the quality impact of packaging processes, and the correlation stability characteristics of historical qualified packaging processes. It also dynamically adjusts the node correlation weights according to the current packaging process stage, so that the packaging quality correlation structure can be updated with the changes in forming, filling, sealing, and conveying processes. This avoids the problem of weakening the key quality impact relationships under different processes due to the use of fixed sensor correlations.

[0020] This invention introduces packaging process constraints into the GDN graph deviation network, filters effective adjacent nodes by the current packaging process stage, and aggregates graph attention features by combining dynamic node association weights. This enables the model to not only utilize the temporal state of the target sensor node itself, but also comprehensively reflect the state of associated nodes that have an actual impact on packaging quality, thereby improving the ability to identify multi-sensor collaborative anomalies and local process anomalies.

[0021] This invention simultaneously constructs sensor state deviation and sensor correlation deviation. Sensor state deviation is used to characterize the change in the actual monitoring state relative to the normal quality prediction state, and sensor correlation deviation is used to characterize the change in the correlation relationship of the current node relative to the normal packaging correlation benchmark. This enables the identification of packaging quality problems where the value of a single sensor has not yet shown obvious abnormalities, but the collaborative relationship between multiple sensors has already shifted, reducing the risk of missed detection caused by relying solely on a single-point threshold or a single prediction error.

[0022] This invention can also determine abnormal sensor nodes based on the characteristics of packaging bag quality deviation, and further filter related abnormal nodes along the packaging quality dynamic correlation diagram, mapping the abnormal state to specific packaging processes and abnormal action locations, so that the final monitoring results can not only give the quality status of the packaging bag, but also reflect the type of quality abnormality and the area of ​​abnormal action, providing clearer data basis for adjusting packaging equipment parameters, investigating abnormal processes, and tracing packaging quality. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent monitoring method for feed packaging bag quality based on sensor networks proposed in this invention. Figure 2 This is a flowchart illustrating the construction of a dynamic correlation graph for the packaging quality of the intelligent monitoring method and system for feed packaging bag quality based on sensor networks proposed in this invention. Figure 3 This is a schematic diagram of the dynamic correlation distribution and projection of sensors under the packaging process constraints of the intelligent monitoring method and system for feed packaging bag quality based on sensor networks proposed in this invention. Figure 4 This is a schematic diagram of the GDN graph deviation network for identifying the quality deviation of feed packaging bags, which is part of the intelligent monitoring method and system for feed packaging bag quality based on sensor networks proposed in this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0025] refer to Figures 1-4 A method and system for intelligent monitoring of feed packaging bag quality based on sensor networks includes the following steps: S1 collects multi-source quality monitoring data and packaging equipment operation status data during the feed packaging process, and forms packaging quality time-series data according to packaging bag identification, node identification and collection time. S2, perform time alignment and process division on the packaging quality time series data, and encode the sensor data of each process stage to obtain the packaging quality state sequence; S3, acquire sensor node attributes and historical monitoring data of qualified packaging, establish the quality impact relationship of packaging process, construct a dynamic correlation diagram of packaging quality and continuously update it to obtain correlation change records; S4. Input the packaging quality status sequence and packaging quality dynamic correlation diagram into the GDN graph deviation network, and obtain the quality status prediction result through process constraint graph attention aggregation; S5. Establish a normal packaging correlation benchmark based on the historical monitoring data of qualified packaging, and calculate the dual deviation by combining the quality status prediction results and correlation change records to obtain the packaging bag quality deviation characteristics. S6. Identify abnormal sensor nodes and associated abnormal nodes based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, and determine the type of quality abnormality and the area of ​​abnormality. S7. Determine the quality status of the packaging bag based on the characteristics of the packaging bag quality deviation, and generate the packaging bag quality monitoring results by combining the identified quality anomaly types and the areas affected by the anomalies.

[0026] In this embodiment, S1 includes: Packaging bags entering the feed packaging equipment are identified with labels to distinguish monitoring data corresponding to different packaging bags on the continuous production line. From the moment the packaging bag enters the forming area until it is filled, sealed, and leaves the conveying area, the monitoring data generated by the same packaging bag is associated with the corresponding packaging bag label.

[0027] Multi-source quality monitoring data is continuously collected through a sensor network, while packaging equipment operating status data is also read. The multi-source quality monitoring data includes data reflecting the quality status of the packaging bags, such as bag position, tension, weight, temperature, pressure, deformation, and vibration; the packaging equipment operating status data includes the operational status of the traction mechanism, loading mechanism, sealing mechanism, and conveying mechanism.

[0028] For continuous numerical sensors, record the node identifier, packaging bag identifier, acquisition time, and corresponding monitored value. For equipment operation status data, record the corresponding equipment action status and the time of occurrence, so that the sensor monitored values ​​can establish a time correspondence with the actual actions of the packaging equipment.

[0029] Different sensor nodes can have different sampling frequencies. For continuously changing monitored quantities such as temperature, pressure, weight, and tension, continuous sampling data within the current packaging cycle is retained; for state-type data such as equipment start-up, shutdown, and mechanism operation, the state change time and state duration range are retained, and it is not required that the original sampling times of different nodes be completely consistent.

[0030] Multi-source quality monitoring data and packaging equipment operating status data belonging to the same packaging bag are organized according to the collection time to form packaging quality time-series data. The packaging quality time-series data retains the packaging bag identification, node identification, and collection time, providing a basis for subsequent determination of the packaging bag, sensor source, and time location corresponding to each monitoring data.

[0031] During continuous packaging, when the previous packaging bag is still in the conveying stage and the next packaging bag has entered the forming stage, the monitoring data attribution relationship is established according to the packaging bag identification and the corresponding equipment operation position to avoid the mixing of sensor data of adjacent packaging bags.

[0032] Packaging quality time-series data serves as input for time alignment, process division, and node status analysis, and does not directly determine the final packaging bag quality status based on raw sensor values.

[0033] In this embodiment, S2 includes: Time-series data on packaging quality is aligned and divided into packaging processes, establishing a correspondence between the monitoring status of different sensor nodes in the same packaging bag and the same packaging process.

[0034] First, a unified time series is established based on the data acquisition time of each sensor. The unified time series covers the time range from when a single packaging bag enters the forming area to when it is completed being conveyed, and the corresponding time position is determined according to the sensor sampling cycle and the working cycle of the packaging equipment.

[0035] For continuous numerical monitoring data, when a corresponding sample value exists at a target time position in a unified time series, a direct correspondence is established; when the target time position lies between adjacent sample positions, the corresponding state is determined based on the time interval and change relationship between adjacent monitoring values. During online monitoring, only currently acquired monitoring data is used; future monitoring values ​​that have not yet arrived are not used to infer the current node's state.

[0036] For time locations lacking valid sampling data, the corresponding data validity information is retained, and the missing states are not directly used as normal monitoring values ​​in subsequent analyses. For status-based data such as the start-up of the loading mechanism, the closure of the sealing mechanism, and the operation of the conveying mechanism, a time correspondence is established based on the start time and duration of the status.

[0037] After time alignment is completed, the packaging process stages are determined based on the operating status of the packaging equipment and changes in sensor responses. The forming stage is determined when the bag traction or clamping mechanism enters operation and the bag's position, tension, or deformation nodes show continuous responses; the filling stage is determined when the filling mechanism is operating and the weight node shows a continuous increase; the sealing stage is determined when the sealing mechanism enters the pressing state and the temperature, pressure, or sealing deformation nodes show corresponding responses; and the conveying stage is determined when the packaging bag enters the conveying mechanism and the vibration, position, or attitude nodes show corresponding changes.

[0038] The packaging process utilizes a combination of equipment operating status and sensor response for verification. When the equipment's action signal has switched but the sensor has not yet generated a valid response, the process boundary is determined by combining the actual response changes, thereby reducing process division deviations caused by mechanical action delays or sensor response delays.

[0039] Based on the time range corresponding to each packaging process stage, sensor data belonging to the same packaging bag and the same process stage are extracted from a unified time series to form process stage monitoring data. Subsequently, the process stage identifier and the corresponding monitoring data are jointly coded and arranged according to the chronological relationship of forming, filling, sealing and conveying to obtain the packaging quality status sequence.

[0040] The packaging quality status sequence retains the correspondence between sensor node identifiers, monitoring time, node monitoring status, and packaging process stages, enabling subsequent models to distinguish the quality meaning of the same sensor status change under different packaging processes.

[0041] For historical monitoring data of qualified packaging bags, a consistent time alignment and process division method is adopted to form a historical packaging quality status sequence, which is used for subsequent quality correlation analysis, GDN model training and normal packaging correlation benchmark establishment.

[0042] In this embodiment, S3 includes: Acquire sensor node attributes and historical monitoring data of qualified packaging, establish the quality influence relationship of packaging process based on the packaging quality status sequence, construct a dynamic correlation diagram of packaging quality that can change with the packaging process stage, and form a correlation change record.

[0043] Historical monitoring data for qualified packaging comes from packaging bags that have passed quality inspection and been confirmed to meet packaging requirements. This historical monitoring data is linked to the corresponding packaging bag identification and quality inspection results, including sensor monitoring status and packaging equipment operating status. Data for packaging bags that fail quality inspection are not included in the establishment of normal quality correlations.

[0044] First, the sensor type and installation location of each sensor node are obtained from the node configuration record, and then node attribute features are formed by combining them with the current packaging process stage. Node attribute features are used to distinguish the monitored objects of the nodes and their corresponding work positions, enabling the model to identify the physical properties and process functions of different nodes such as temperature, pressure, weight, and tension.

[0045] The temporal change characteristics of each node within the current monitoring time window are extracted from the packaging quality status sequence. These characteristics include the node's current state, state changes at adjacent time points, direction of continuous change, degree of local fluctuation, and state duration. The monitoring time window is determined based on the sensor sampling period and the duration of the packaging process, ensuring that the window can retain the main node changes in the current process.

[0046] For the time-series variation characteristics of different physical quantities, scaling is performed based on the normal variation range corresponding to the same sensor type and the same process in the historical monitoring data of qualified packaging to reduce the impact of unit differences of different physical quantities such as temperature, pressure and weight on node correlation analysis.

[0047] By fusing node attribute features with node temporal variation features, sensor node features are obtained. These features simultaneously include the node's physical attributes and current packaging state, enabling the same node to represent different states under different packaging processes or different packaging bag conditions.

[0048] Further, the correlation of node state changes is determined based on sensor node characteristics. For two nodes, their state changes within the same monitoring time window are correlated, comparing the direction of change, the timing of major changes, continuous changes, and state response relationships. For nodes with normal sequential response relationships, the response time difference in historical qualified packaging data is used for correlation matching to avoid misjudging a lag in normal process response as a weakening correlation.

[0049] When two nodes only show accidental synchronous changes at a single time location, and lack a stable correspondence within a continuous monitoring time window, their state changes are less correlated; when two nodes maintain synchronous changes or stable sequential responses within a continuous time range, their state changes are more correlated.

[0050] Simultaneously, a quality impact relationship for the packaging process is established based on sensor node attributes and packaging process stages. This relationship indicates whether the physical states monitored by different sensor nodes jointly participate in the current packaging quality formation process. It is determined based on sensor type, installation location, and packaging equipment process structure, and is not solely based on data correlation.

[0051] During the forming stage, the bag's position, tension, and deformation nodes jointly reflect the bag's unfolding and clamping state, thus affecting the forming quality. During the filling stage, the weight and bag's stress nodes jointly reflect the filling amount and the bag's load-bearing state, thus affecting the filling quality. During the sealing stage, temperature, pressure, sealing position, and deformation nodes in the sealing area jointly participate in the sealing quality formation, thus affecting the sealing quality. During the conveying stage, vibration, position, and attitude nodes jointly reflect the bag's conveying stability, thus affecting the conveying quality.

[0052] For cross-process node relationships, a corresponding relationship is established based on the sequence of processes and the actual impact on quality. For example, a deviation in the molding position may affect the sealing position, but its process impact relationship is different from the direct correlation between sealing temperature and sealing pressure.

[0053] By combining the correlation between node state changes and their impact on packaging process quality, node combinations with actual quality impact and corresponding state changes are selected to form an initial quality correlation set. Node combinations lacking quality effects and without corresponding stable states are not included in the initial quality correlation set.

[0054] like Figure 2 As shown, after forming the initial quality association set, the historical monitoring data of qualified packaging is further used to determine the association stability and reliability of each node combination.

[0055] Time alignment and process division are performed on the historical monitoring data of qualified packaging. The historical state change relationship of the corresponding node is extracted according to the node combination in the initial quality association set to form historical node association data.

[0056] Based on historical node association data, the degree of common change, the degree of consistency in change direction, and the degree of association persistence of node combinations are calculated to obtain historical association stability characteristics. The degree of common change is used to represent the stability of two nodes undergoing common state changes during normal packaging; the degree of consistency in change direction is used to represent whether nodes maintain the same or stable change direction over a long period; the degree of association persistence is used to represent whether the node relationship can be maintained in a complete process stage or during continuous qualified packaging.

[0057] For node relationships that persist throughout multiple qualified packaging processes, their historical correlation stability is relatively high; for node relationships that are formed only occasionally at a few times and locations, their historical correlation stability is relatively low.

[0058] By jointly encoding the correlation degree of node state changes, the impact of packaging process quality, and historical correlation stability characteristics, a quality correlation reliability feature is obtained, which is used to determine whether the current node combination simultaneously has the actual process action relationship, the current state correspondence relationship, and historical normal stability.

[0059] When a node combination has a high degree of correlation with the current state, a direct impact on process quality, and a stable historical association, increase the reliability of its quality association. When a node combination only has short-term data correlation but lacks actual quality impact or has an unstable historical association, decrease the reliability of its quality association. When a node combination has a clear impact on process quality and high historical stability, but its current correlation decreases due to short-term disturbances, retain the corresponding relationship by combining the stable characteristics of historical associations to avoid instantaneous noise directly destroying the normal graph structure.

[0060] Trustworthy quality association relationships are screened based on the reliability characteristics of quality associations and the current packaging process stage. First, the effective sensor nodes in the current process are identified, and node combinations containing effective nodes are extracted from the initial quality association set. Node relationships that directly participate in the formation of quality in the current process and have a high degree of reliability are retained. Whether to retain auxiliary relationships is determined based on their degree of reliability. Node combinations that do not match the current process are excluded from the current effective adjacency relationships.

[0061] The association weights of reliable quality associations are determined based on the correlation of node state changes, the stability of historical associations, and the impact of packaging process quality. This information is processed using a standardized scale before being used in the weight calculation. The higher the correlation of the current state, the more stable the historical association, and the more direct the effect of the process, the greater the association weight of the corresponding node combination.

[0062] One feasible weighting method involves weighting the normalized node relevance and historical correlation stability according to their respective weights, and then multiplying this weighted combination by the current process quality impact coefficient. For direct quality impact relationships, a higher process impact coefficient is used; for auxiliary quality impact relationships, a lower coefficient is used; and relationships that do not match the current process are excluded. Each combination coefficient is determined through the correlation stability in qualified packaging data, and the weighting calculation rules within the same process are kept consistent.

[0063] Using effective sensor nodes as graph nodes, reliable quality correlations as graph edges, and corresponding correlation weights as graph edge attributes, a dynamic correlation graph of packaging quality corresponding to the current packaging process is constructed. The node identifiers in the graph structure are consistent with those in the sensor network, and the graph edges record node combinations and correlation weights, and are associated with the current packaging process stage and monitoring time.

[0064] When the packaging process switches from filling to sealing, the correlation between weight and bag stress points can be reduced, while the correlation between temperature, pressure, and sealing deformation points increases accordingly. For node combinations that still have actual quality effects across processes, the corresponding relationships are retained based on the reliable characteristics of quality correlation, thereby avoiding the use of fixed correlation diagrams to describe the complete packaging process.

[0065] During continuous monitoring, the dynamic correlation diagram of packaging quality is updated according to the monitoring time window. The graph structure of adjacent monitoring time positions is compared, and changes in effective correlation nodes, the addition or disappearance of graph edges, and changes in the correlation weight of the same node combination are recorded. The above changes are correlated with node identifiers, process stages, and monitoring time to form a correlation change record.

[0066] Related change records are saved chronologically to determine whether changes in node relationships are short-term disturbances or persistent anomalies. For graph structure changes caused by normal process switching, the process stage identifier is retained and compared with the normal packaging correlation benchmark of the same process in subsequent analysis; relationship changes caused by process switching are not directly identified as quality anomalies.

[0067] like Figure 3 As shown, Figure 3 This is a schematic diagram of the dynamic correlation distribution and projection of sensors under the constraints of the packaging process.

[0068] Figure 3 The initial quality association distribution on the left represents the node association state formed by the initial quality association set before reliable association screening and weight adjustment for the current packaging process. The horizontal and vertical positions in the figure correspond to different sensor nodes or node combinations, and the local undulations of the surface represent the strength of candidate associations between node combinations.

[0069] Since the initial quality association set contains multiple quality associations that may exist in different packaging processes, the distribution on the left shows multiple local association regions. Figure 3 The sensor node plane on the lower left is used to represent the node space participating in the quality correlation analysis. The projection relationship between the upper distribution and the lower node plane represents the correspondence between each correlation region and a specific node combination.

[0070] After process constraint correction, valid sensor nodes are determined based on the current packaging process. Reliable quality association relationships are then screened using quality association reliability features, and the corresponding association weights are adjusted. Node relationships directly related to the current process are strengthened, while mismatched relationships are weakened or eliminated.

[0071] Figure 3The normalized dynamic correlation distribution on the right represents the current quality correlation state after the completion of process constraints, with the main correlation areas related to the current process being more prominent. The correlation projection diagram at the bottom right represents the projection of the dynamic correlation distribution into the sensor node space. The shades of gray correspond to the correlation strength of the node combinations, with darker areas representing node relationships with relatively high correlation under the current process.

[0072] Figure 3 The curved surfaces and projections in the graph are used to visually illustrate the changing mechanism of node relationships. In actual processing, the graph structure data is still composed of node identifiers, reliable quality relationships, association weights, and monitoring time. It is not required to convert sensor relationships into images before performing GDN calculations.

[0073] In this embodiment, S4 includes: The packaging quality status sequence and packaging quality dynamic correlation diagram are input into the GDN graph deviation network, and the quality status prediction result is obtained through process constraint graph attention aggregation.

[0074] Before online prediction, a GDN (Graphical Deviation Network) is trained using historical monitoring data of qualified packaging. First, a historical packaging quality status sequence is formed using time alignment and process coding methods consistent with the current packaging quality time series data. Then, a corresponding historical packaging quality dynamic association graph is established according to the graph construction method in S3, so that the training phase and the online prediction phase use consistent node definitions, association rules, and process constraints.

[0075] Training and validation samples are divided according to the packaging bag markings to prevent continuous monitoring data from the same packaging bag from being simultaneously included in model training and independent validation processes. Training samples are used to optimize GDN model parameters, while validation samples are used to check normal prediction errors and provide data for determining the normal deviation range of subsequent nodes.

[0076] For each training sample, the continuous node states before the target monitoring time are extracted from the historical packaging quality state sequence as the model input, and the actual normal node states corresponding to the target monitoring time are used as the prediction target.

[0077] When constructing training inputs, only the monitoring status, process information and corresponding dynamic correlation diagrams obtained before the target monitoring time are used. The actual monitoring values ​​at the target time location are not used as prediction inputs to avoid the actual monitoring information from entering the prediction process in advance.

[0078] The GDN (Graph Deviation Network) uses sensor node embedding to represent node attributes, represents node quality relationships through a graph structure, and aggregates associated node information using a graph attention mechanism. This implementation uses the effective adjacency relationships in the dynamic packaging quality correlation graph as adjacency constraints for graph attention processing. It incorporates node association weights and the current packaging process stage into the attention calculation, ensuring that the selection and contribution of adjacent nodes are constrained by the quality relationships of the current process.

[0079] During model training, the prediction loss is formed based on the difference between the predicted node state and the actual node state of historical qualified packaging bags. The prediction error of the normal state is reduced by updating parameters, so that the model can learn the laws of node state changes and collaborative changes between nodes during normal packaging.

[0080] The prediction loss can be composed of the prediction errors of each effective node, and normalized according to the normal scale corresponding to different sensors to avoid training bias caused by differences in the numerical range of temperature, pressure and weight.

[0081] After training, the normal prediction error under different packaging processes is checked using the validation samples, and the node embedding, graph attention, process encoding and state prediction parameters are saved to obtain the trained GDN graph bias network.

[0082] During online prediction, the temporal state of nodes before the target monitoring time is extracted from the packaging quality status sequence, and the dynamic correlation graph of packaging quality constructed in the corresponding time window is read to determine the effective adjacent nodes and correlation weights of the target sensor node.

[0083] For a target sensor node, its temporal state is matched with the temporal states of its effective neighboring nodes to form an attention response feature. The attention response feature represents the degree of matching between the target node and different neighboring nodes in the current state change; the node association weight represents the degree of quality association of the corresponding node combination in the current process; the packaging process quality influence relationship is used to determine whether the node combination is consistent with the current quality formation process.

[0084] Attention response features, node association weights, and the quality impact relationship of the current process are integrated to form process-constrained attention weights. For adjacent nodes with high quality association credibility in the current process and a stable state correspondence with the target node, their aggregation contribution is increased; for adjacent responses that are short-term relevant but do not match the quality impact relationship of the current process, they are suppressed.

[0085] The attention responses of retained effective neighboring nodes are normalized to ensure a consistent computational scale for their contributions during the aggregation process of the same target node. When there are no other effective neighboring nodes that meet the conditions in the current process, the target node's own temporal state is retained for prediction, and mismatched nodes from other processes are not forcibly introduced.

[0086] The temporal states of effective adjacent nodes are weighted and aggregated based on the attention weights of process constraints to obtain the associated quality aggregation feature of the target node. This feature represents the combined effect of other effective nodes in the current packaging process on the normal quality state of the target node.

[0087] The target node's own temporal state, associated quality aggregation features, and current process stage are fused together. The normal node state corresponding to the target monitoring time is predicted by the trained GDN graph deviation network, and the quality state prediction result is obtained.

[0088] A correspondence is established between the quality status prediction results and the target node identifier and target monitoring time. Other valid nodes are processed in the same way to form the quality status prediction results for each sensor node of the current packaging bag.

[0089] Once the actual data for the target monitoring time arrives, the current packaging quality dynamic correlation diagram is updated using the corresponding actual monitoring data, and the newly formed correlation status is added to the correlation change record. The historical dynamic correlation diagram used for target time prediction and the current correlation status updated after the actual data arrives retain time information respectively. The former participates in normal state prediction, and the latter participates in correlation deviation analysis, avoiding the prediction process where the actual data for the target time enters the prediction process at the same time position in advance.

[0090] In this embodiment, S5 includes: A normal packaging correlation benchmark is established based on historical monitoring data of qualified packaging, and sensor state deviation and sensor correlation deviation are generated respectively. The packaging bag quality deviation characteristics are obtained through joint coding.

[0091] The establishment of normal packaging correlation benchmarks adopts dynamic correlation with current packaging quality. Figure 1 The node attributes, process division, reliable association screening, and association weight formation methods are consistent.

[0092] First, the historical monitoring data of qualified packaging is aligned by time and divided into processes to form a historical packaging quality status sequence. Then, a dynamic correlation diagram of historical packaging quality is constructed according to the dynamic mapping method in S3.

[0093] The historical packaging quality dynamic correlation diagram is classified according to the stages of forming, filling, sealing and conveying processes. The correspondence between the same node combinations is established according to the node identifier. The effective adjacency relationship, correlation weight change range and correlation continuity of each node combination in the normal packaging process are extracted to form the normal packaging correlation benchmark.

[0094] The normal packaging association benchmark includes packaging process stages, node combinations, normal valid adjacency relationships, the range of association weight changes, and association persistence characteristics. For node relationships that do not normally exist in historical qualified packaging processes, the corresponding ineffective association state is retained to identify newly added abnormal associations.

[0095] Normal packaging relationship benchmarks are established separately for each packaging process stage, and the node relationships of different processes are not mixed into a single fixed benchmark. The node definitions, process divisions, and relationship weight expressions are consistent with the current dynamic relationships. Figure 1 This allows the two to be directly compared based on the same node combinations and process stages.

[0096] After establishing a normal packaging correlation benchmark, the predicted quality status is compared with the actual monitoring status corresponding to the target monitoring time. For any node, the predicted normal status and the actual monitoring status are obtained based on the node identifier and the target monitoring time, and the difference in node status prediction is calculated.

[0097] For numerical monitoring data, the difference between the predicted value and the actual monitored value is used to describe the prediction error, and the data is normalized according to the normal change scale corresponding to the same sensor and the same packaging process to form the sensor state deviation.

[0098] For a node with multiple monitoring dimensions, the prediction difference of each dimension is calculated separately, and then the state relationship between the dimensions is jointly processed to form the sensor state deviation corresponding to the node.

[0099] Sensor state deviation is used to represent the degree of deviation of a node's actual state from its normal predicted state. For short-term differences that occur only at a single monitoring time location, the deviation and time information are retained, and its changes are observed within a continuous monitoring time window, without directly identifying it as a continuous anomaly.

[0100] At the same time, the node association status is extracted from the current packaging quality dynamic association diagram, and the corresponding normal packaging association benchmark is read according to the packaging process stage and node combination.

[0101] Compare the current association status with the effective adjacency relationships and association weight changes in the normal packaging association benchmark. When a stable association relationship that existed in the historical qualified packaging process is significantly weakened in the current packaging bag, a change in association strength is formed; when a normal effective adjacency node disappears, a change in association node is formed; when a new association that does not exist in the normal packaging association benchmark appears, a new association change is formed.

[0102] Further reading of the correlation change records formed by S3, and matching the current correlation difference with the changes in adjacent monitoring time windows, determines whether the abnormal correlation persists.

[0103] For correlation changes that occur only at a single time location and subsequently return to normal, they are retained as short-term correlation fluctuations; for situations where multiple consecutive monitoring time locations deviate from the normal range, the correlation strength continues to change, or abnormal adjacency relationships persist, the corresponding correlation anomaly level is increased.

[0104] Sensor correlation deviation is formed based on changes in correlation strength, changes in effective correlation nodes, and continuous changes in correlation. Sensor correlation deviation takes node combinations as the basic object, retaining the node identifiers at both ends of the correlation, the corresponding process stage, and the time of anomaly occurrence.

[0105] Sensor state deviation and sensor association deviation describe changes in the node's own state and changes in the node's cooperative relationship, respectively. Both types of deviations are not required to increase simultaneously.

[0106] For example, during the sealing process, the actual monitored values ​​of the temperature and pressure nodes may be close to the GDN prediction results, but under normal packaging conditions, they have a stable cooperative relationship with the sealing deformation node. When this relationship continuously weakens during the current packaging process, even if the deviation of a single node is small, the abnormal sealing quality relationship can be reflected by the sensor correlation deviation.

[0107] Therefore, for purely correlated anomalies that persist and can correspond to a clear impact on the quality of the packaging process, the corresponding nodes and related anomaly information should be retained, and they should not be directly excluded just because the node's own value is still within the normal range.

[0108] The sensor state deviation and sensor association deviation are jointly encoded according to the node identifier, the correspondence of associated nodes and the monitoring time to obtain the packaging bag quality deviation characteristics.

[0109] The packaging bag quality deviation characteristics retain the degree of node state deviation, the degree of correlation deviation, the correspondence between related nodes, the direction of change and the duration. Instead of merging all node deviations into a single packaging bag score and deleting the node source, it provides complete data for subsequent abnormal node screening and quality anomaly type judgment.

[0110] like Figure 4 As shown, Figure 4 This is a schematic diagram of packaging bag quality deviation identification using the GDN graph deviation network.

[0111] Figure 4 The predicted state distribution at the top left represents the normal node state predicted by GDN based on the packaging quality state sequence and the packaging quality dynamic correlation diagram, while the actual monitored state distribution at the bottom represents the actual node state obtained by the sensor network at the same monitoring time and location.

[0112] During normal packaging, the predicted state and the actual monitored state have similar trends. When an anomaly occurs at a local node, the actual monitored state will experience a local rise, fall, or shape shift relative to the predicted state, resulting in a sensor state deviation.

[0113] In addition to the differences in the state of the nodes themselves, the correlation strength, effective adjacency relationship and continuous change differences between the current packaging quality dynamic correlation diagram and the normal packaging correlation benchmark form the sensor correlation deviation. Figure 4 The dual-bias mapping in the middle represents mapping two types of biases to the corresponding sensor node space according to node identifiers and association relationships.

[0114] Figure 4 The abnormal nodes in the right-hand node region represent abnormal sensor nodes identified through state deviation or correlation deviation. Nodes that maintain a reliable quality correlation with these nodes and exhibit synchronous abnormal changes are considered as associated abnormal nodes. Nodes with small state deviations but whose correlation relationships are consistently abnormal retain their candidate status as associated abnormal nodes.

[0115] The abnormal projection area in the lower right corner represents the concentrated state of the dual deviations in the sensor node space. Darker gray areas correspond to node relationship areas where the deviations are more concentrated. This projection is used to visually illustrate the relationship between dual deviations and abnormal nodes. In actual monitoring, node status data, graph correlation, and deviation characteristics are still used for calculation, and the quality status is not directly determined by the image grayscale value.

[0116] In this embodiment, S6 includes: Based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, abnormal sensor nodes and associated abnormal nodes are identified, and when an abnormality exists, the type of quality abnormality and the area of ​​abnormality are determined.

[0117] First, the trained GDN graph bias network is used to predict the normal state of the verification samples in the historical monitoring data of qualified packaging. The prediction results of the verification samples are compared with the actual state corresponding to the target monitoring time to obtain the node prediction error under normal packaging conditions.

[0118] Prediction errors are categorized according to sensor node type and packaging process stage. The distribution range of normal prediction errors during qualified packaging is statistically analyzed to form the normal deviation range of nodes. The normal deviation range of nodes is derived from the prediction error of GDN for qualified calibration samples, rather than directly using the normal range of the original sensor monitoring values, so that it has the same comparison basis as the sensor state deviation in S5.

[0119] For sensor correlation deviation, the normal range of variation of correlation strength and effective correlation relationship under the same process is determined based on the normal packaging correlation benchmark, and the correlation anomaly judgment is formed by combining the correlation continuity status.

[0120] Based on the characteristics of packaging bag quality deviation, determine whether the node status deviation and node correlation deviation exceed the corresponding normal change range, and filter abnormal candidates based on the deviation persistence.

[0121] When a node's state deviation consistently exceeds the corresponding normal deviation range, it is identified as a candidate node for state anomaly. When a node's own state deviation is small, but a certain quality association it participates in consistently deviates from the normal packaging association benchmark, the endpoint node of that association is included in the association anomaly candidate, and the corresponding abnormal association source is retained.

[0122] For node relationships that only experience short-term changes and subsequently return to normal, the degree of anomaly determination is reduced based on the relationship change record; for relationships that continuously deviate from the normal packaging relationship benchmark, the corresponding relationship anomaly status is retained.

[0123] By jointly analyzing the nodes with continuous abnormal status and the nodes with continuous associated abnormal relationships, abnormal sensor nodes are selected based on the degree of deviation, continuous status and the impact on the quality of the packaging process.

[0124] Therefore, abnormal sensor nodes can be either nodes whose own state is continuously abnormal, or nodes involved in a continuous abnormal quality correlation. For purely correlated anomalies, a sensor device failure is not directly identified in a particular node; instead, the abnormal correlation and its two endpoints are retained and used as a combination of packaging quality abnormal nodes in the analysis.

[0125] After identifying the abnormal sensor nodes, read the nodes with reliable quality correlation with them along the packaging quality dynamic correlation graph, and filter the associated abnormal nodes based on the node correlation weight, the duration of the correlation anomaly, and the synchronous change relationship of the node quality deviation.

[0126] When a related node has a strong current quality correlation with an abnormal sensor node, and the two form a stable corresponding deviation change within a continuous monitoring time window, it is identified as a related abnormal node. For purely related anomalies, the node correspondence is retained according to the two endpoints of the anomaly relationship, so that anomalies formed solely by changes in the cooperative relationship can still be included in subsequent localization.

[0127] For cases where node state deviations are large but other trusted associated nodes lack synchronous changes, retain single-node anomaly information and do not force the generation of non-existent associated anomaly nodes.

[0128] After completing the identification of abnormal nodes, the abnormal sensor nodes and the sensor types, installation locations and packaging process stages corresponding to the associated abnormal nodes are read. Based on the node configuration records and the process structure of the packaging equipment, the correspondence between the sensor installation locations and the functional areas of the packaging bags is established.

[0129] For sensors installed on actuators, the location of any abnormality in the packaging bag is determined based on the corresponding actuator's position. For example, a sealing pressure detection node installed on a sealing actuator maps any quality abnormality to the sealing area of ​​the packaging bag; a tension detection node installed on a traction mechanism corresponds to any abnormality in the bag's unfolding and forming area.

[0130] For weight and bag stress nodes, anomalies are mapped to the loading and load-bearing areas; for vibration, position, and attitude nodes, anomalies are mapped to the corresponding bag action areas during the conveying stage, based on the position of the conveying mechanism and the current state of the packaging bag.

[0131] Further, the type of quality anomaly is determined based on the combination of abnormal nodes and their impact on the quality of the packaging process. When the abnormal node mainly involves tension, position, and deformation and occurs during the forming stage, it is determined to be an abnormal bag forming state; when the abnormal node mainly involves weight and bag stress and occurs during the filling stage, it is determined to be an abnormal filling state; when the abnormal node involves temperature, pressure, sealing position, and sealing deformation and forms a continuous abnormality during the sealing stage, it is determined to be a sealing quality anomaly; when vibration, position, and attitude nodes form a continuous abnormality during the conveying stage, it is determined to be a conveying stability anomaly.

[0132] The anomaly type is determined by the combination of anomaly nodes, the packaging process stage, and the actual impact on quality, and is not determined directly based on the name of a single sensor node.

[0133] For cases where the same combination of abnormal nodes may correspond to multiple abnormal types, the corresponding abnormal type is determined by combining the packaging process where the abnormality first occurs continuously, the main deviation nodes, and the impact of process quality, while retaining other related abnormal information.

[0134] When it can only be determined that an abnormality has occurred in a certain packaging process, but the sensor information is insufficient to confirm the specific physical defect, the corresponding process quality abnormality type is output, and the indirect state change is not directly identified as the already confirmed packaging bag damage or leakage.

[0135] An abnormal action area is formed based on the quality anomaly type and the correspondence between the sensor and the action area of ​​the packaging bag. If no node or correlation meeting the anomaly conditions is detected, no quality anomaly type or abnormal action area is generated, and the packaging bag quality status judgment continues.

[0136] In this embodiment, S7 includes: The quality status of the packaging bags is determined based on the characteristics of the packaging bag quality deviations, and the packaging bag quality monitoring results are generated by combining the identified quality anomaly types and the areas affected by the anomalies.

[0137] The quality status of packaging bags includes normal status, quality concern status, and quality abnormal status, which are judged based on the deviation of node status, the deviation of node correlation, and the duration of abnormality.

[0138] When the deviation of each node's status is within the corresponding normal deviation range, the node association relationship is consistent with the normal packaging association benchmark, and no continuous abnormal nodes or abnormal association relationships are formed, the corresponding packaging bag is determined to be in a normal state.

[0139] When some nodes exhibit status deviations outside the normal range, but the degree of persistence is insufficient, or the node correlation has not yet formed a sustained abnormality, the corresponding packaging bag is designated as a quality concern status. When the correlation shows certain abnormal changes, but it is not yet possible to clearly identify the specific process quality abnormality, it is also designated as a quality concern status, and the relevant node and time information are retained.

[0140] When sensor status deviation persists and corresponds to a clear process quality anomaly, the quality anomaly status is determined by combining the abnormal node and its correlation. Similarly, if a node's own status deviation is small, but its correlation deviation consistently exceeds the normal correlation variation range, and can be pinpointed to a specific packaging process, it can also be determined as a quality anomaly status.

[0141] Therefore, the only condition for determining a quality anomaly is not the simultaneous existence of state deviation and associated deviation. Instead, it is determined by comprehensively considering the degree, duration, and impact of the two types of deviation on process quality, thus avoiding overlooking packaging quality issues that only manifest as abnormal sensor coordination.

[0142] For packaging bags with missing sensor data or insufficient effective monitoring time, making it impossible to reliably determine their quality status, they are not directly classified as being in a normal state. Instead, the data validity information is retained, they are included in the quality concern status, and corresponding verification information is provided.

[0143] After determining the quality status of the packaging bags, establish a corresponding relationship between the packaging bag identification, quality status, abnormal nodes, associated abnormal nodes, quality abnormality types, and abnormal impact areas to form the packaging bag quality monitoring results.

[0144] For packaging bags in normal condition, the monitoring results record the packaging bag identification and normal status, without generating any non-existent abnormal types or abnormal areas. For packaging bags in quality concern condition, the monitoring results record deviation nodes, abnormal correlations, deviation duration, and related process information. For packaging bags in quality abnormal condition, the monitoring results record abnormal nodes, associated abnormal nodes, quality abnormality types, and abnormal areas, which are used for subsequent abnormal packaging bag sorting, equipment inspection, and quality traceability.

[0145] The results of packaging bag quality monitoring can be saved to the packaging production quality database or output to the packaging equipment control terminal or quality management terminal. When multiple packaging bags exhibit similar abnormal node combinations in the same process, the corresponding equipment operation position can be determined based on the saved monitoring results, providing a basis for equipment inspection and process parameter adjustment.

[0146] This embodiment also provides a sensor network-based intelligent monitoring system for feed packaging bag quality, used to execute the above-mentioned intelligent quality monitoring method, including a data acquisition module, a state sequence construction module, a dynamic correlation graph construction module, a GDN quality prediction module, a quality deviation analysis module, an anomaly location module, and a quality result generation module.

[0147] The data acquisition module is connected to sensor nodes and equipment operation status interfaces deployed at different locations on the feed packaging equipment. It is used to collect multi-source quality monitoring data and packaging equipment operation status data, form packaging quality time series data according to packaging bag identification, node identification and collection time, and retain sensor node configuration records.

[0148] The state sequence construction module is used to align the packaging quality time series data in time. It determines the packaging process stage based on the operating status of the packaging equipment and the changes in sensor response. It then co-encodes the corresponding sensor status and process stage information to obtain the packaging quality state sequence.

[0149] The dynamic correlation graph construction module is used to acquire sensor node attributes and historical monitoring data of qualified packaging, establish the quality influence relationship of packaging process, form an initial quality correlation set based on the correlation degree of node status changes, determine the reliable characteristics of quality correlation by combining historical correlation stability characteristics, and select reliable quality correlation relationships and adjust correlation weights according to the current process to continuously construct a dynamic correlation graph of packaging quality and correlation change records.

[0150] The GDN quality prediction module is used to establish training samples and historical dynamic correlation graphs based on historical monitoring data of qualified packaging, and to train the GDN graph deviation network. During online operation, it determines effective adjacent nodes based on the current packaging quality status sequence and dynamic correlation graph, forms associated quality aggregation features through process constraint graph attention aggregation, predicts the normal node status corresponding to the target monitoring time, and obtains the quality status prediction result.

[0151] The quality deviation analysis module is used to establish a normal packaging correlation benchmark based on historical monitoring data of qualified packaging. It compares the predicted quality status with the actual monitoring status to form the sensor status deviation, and compares the current node correlation status with the normal packaging correlation benchmark, combining this with correlation change records to form the sensor correlation deviation. The two types of deviations are jointly coded, retaining node identifiers, correlation node correspondences, and time information to obtain the packaging bag quality deviation characteristics.

[0152] The anomaly localization module is used to determine the normal deviation range of nodes by using the GDN prediction error of qualified packaging verification samples. It combines the state deviation, correlation deviation and continuous state in the packaging bag quality deviation characteristics to screen abnormal sensor nodes, determines the associated abnormal nodes according to the packaging quality dynamic correlation diagram, and determines the quality anomaly type and abnormal action area by combining the sensor type, installation location, packaging process quality influence relationship and packaging bag action area correspondence.

[0153] The quality result generation module is used to determine the quality status of packaging bags based on their quality deviation characteristics and associate the corresponding quality status with the packaging bag identification. For packaging bags with quality abnormalities, the module generates packaging bag quality monitoring results by combining the identified quality abnormality types and the areas affected by the abnormalities.

[0154] Through the data processing relationships between the above modules, a complete processing process is completed, from multi-source quality monitoring data acquisition, packaging quality status sequence construction, packaging quality dynamic correlation diagram update, GDN normal status prediction, dual deviation analysis, to packaging bag quality status determination and abnormal action area location.

[0155] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous automated feed packaging production line to monitor the quality of the entire process from bag unfolding, quantitative filling, heat sealing to finished product conveying. The production line uses a single packaging bag as the smallest monitoring object, deploying sensor nodes at the bag traction position, weighing position, heat sealing mechanism, and finished product conveying position. Data collected from each node is correlated with the operating status of the packaging equipment over time. During implementation, the quality judgment is not based on whether a single sensor exceeds a fixed value, but rather on the dynamic quality correlation between multiple sensor nodes under changes in the packaging process, to jointly identify the status of the packaging bag.

[0156] The sensor network includes nodes for bag tension, filling weight, heat sealing temperature, heat sealing pressure, sealing area deformation, airtightness, and conveyor vibration. It simultaneously reads the forming, filling, heat sealing, and conveying statuses of the packaging equipment. Each sensor node adds a bag identifier and acquisition time to the collected data, thus forming time-series data on packaging quality corresponding to a single bag. For sampling times from different sensors, a unified time series is used to establish a time correspondence, enabling joint analysis of monitoring data for the same bag within the same packaging process stage.

[0157] First, based on the changes in the operating status of the packaging equipment and the response changes of each sensor node, the packaging quality time-series data is divided into four packaging process stages: forming, loading, sealing, and conveying. For each packaging process stage, sensor data within the corresponding time range is extracted, and the packaging process stage identifier is jointly encoded with the sensor monitoring status to form a packaging quality status sequence according to the sequence of packaging processes. Thus, the same sensor node has different state semantics in different packaging process stages, avoiding the misinterpretation of sensor relationships in the loading and sealing stages using the same association method.

[0158] Subsequently, node attribute features are generated based on the sensor type, installation location, and current packaging process stage of each sensor node. The temporal variation features of the nodes within a continuous monitoring time window are extracted from the packaging quality state sequence. These node attribute features and temporal variation features are then fused to form sensor node features. An initial quality correlation set is formed based on the correlation between state changes of different nodes and their quality impact on the packaging process. Simultaneously, historical monitoring data corresponding to quality-inspected and qualified packaging bags are acquired, and the common degree of change, consistency of change direction, and correlation persistence of each node combination under different packaging process stages are extracted to obtain historical correlation stability features.

[0159] The quality association credibility features of node combinations are formed based on the correlation degree of node state changes, the impact of packaging process quality, and historical association stability characteristics. For the current packaging process stage, only valid sensor nodes involved in the packaging quality formation process of this stage are retained, and node combinations that do not match the current packaging process are filtered out. Furthermore, the association weights are determined according to the correlation degree of the state corresponding to the credibility node relationship, the historical association stability characteristics, and the impact of the current process quality. A dynamic association graph of packaging quality is constructed, with valid sensor nodes as graph nodes, credibility quality association relationships as graph edges, and corresponding association weights as graph edge attributes. As the packaging bag moves from the filling stage to the sealing stage, the valid nodes in the graph, the connection relationships between nodes, and the association weights change synchronously.

[0160] The packaging quality state sequence and packaging quality dynamic correlation graph are input into the GDN graph deviation network. For the target sensor node, effective neighboring nodes are determined from the dynamic adjacency relationship based on the current packaging process stage, and the corresponding node association weights are read. Feature matching is performed on the node temporal states of the target node and each effective neighboring node to obtain attention response features. Then, the attention response features are fused with the node association weights and the quality influence relationship of the current packaging process. Adjacency responses inconsistent with the current packaging process are suppressed to form process constraint attention weights. The node temporal states of each effective neighboring node are weighted and aggregated according to these attention weights to obtain the associated quality aggregated features. These features are then fused with the target node's own temporal state and the current packaging process stage to output the quality state prediction result of the target sensor node.

[0161] In the quality deviation analysis process, the predicted quality status of each sensor node is compared with the actual monitoring status at the same monitoring time and location to form sensor status deviation. Simultaneously, the normal node correlation status corresponding to different packaging process stages is extracted from the historical monitoring data of quality-inspected and qualified packaging bags to form a normal packaging correlation benchmark. The node correlation status in the current packaging quality dynamic correlation diagram is compared with the normal packaging correlation benchmark at the same packaging process stage to form sensor correlation deviation. Sensor status deviation reflects the degree of deviation of the actual state of a single node from the model-predicted normal state, while sensor correlation deviation reflects the change in the quality coordination relationship between nodes relative to the normal packaging process.

[0162] To enable both types of deviations to participate in the quality judgment of packaging bags, the degree, direction of change, and duration of sensor state deviation and sensor correlation deviation within a continuous monitoring time window are jointly coded to form packaging bag quality deviation characteristics. The judgment boundary of quality state is determined based on the deviation distribution and continuous window change relationship corresponding to the historical monitoring data of qualified packaging bags, rather than using fixed empirical thresholds independent of specific equipment. Sensor nodes whose state deviation continuously increases and whose correlation deviation increases synchronously are identified as abnormal sensor nodes; then, correlated abnormal nodes with synchronous deviation changes are screened along the packaging quality dynamic correlation diagram, and the quality abnormality type and abnormal area are determined according to the sensor type, installation location, and packaging process stage of the abnormal node.

[0163] Table 1. Examples of Feed Packaging Bag Quality Monitoring Results

[0164] As shown in Table 1, the state deviations and correlation deviations of B01, B02, B05, and B10 are all within the normal range of variation formed by qualified historical data, and the sensor node relationships of the corresponding packaging processes remain stable. Therefore, the monitoring results are normal. The loading weight node of B03 shows a certain state deviation, but its correlation with the bag tension node remains relatively stable, indicating that the abnormality is mainly manifested as single node or local state fluctuations. The system identifies this as a quality concern state, rather than directly judging it as a quality abnormality.

[0165] B04 exhibited significant state deviations and correlation deviations simultaneously during the sealing stage. The quality correlation between heat-sealing temperature and heat-sealing pressure was noticeably weakened, and this change was consistent with the quality impact during the sealing stage, thus confirming an abnormal heat-sealing quality. B09 similarly showed a simultaneous increase in both types of deviations during the conveying stage. The dynamic correlation diagram further revealed abnormal synergistic changes between the airtightness node and the conveying vibration node, thus pointing the abnormal effect to the bag's sealing state, rather than simply identifying the vibration node as an independent anomaly.

[0166] B07 illustrates the difference between this invention and single-point threshold monitoring methods. The sensor status deviation of this packaging bag is low, and no obvious abnormalities are observed when individually examining nodes such as heat-sealing temperature. However, the correlation deviation between temperature and pressure nodes has increased significantly, indicating a change in the original cooperative relationship between multiple nodes during the sealing process. By comparing the normal packaging correlation benchmark with the current dynamic correlation status, correlation-related sealing anomalies can be identified before individual sensor values ​​show obvious abnormalities, and the packaging bag can be designated as a quality concern.

[0167] B06 and B08 correspond to decreased sealing stability and abnormal forming stress, respectively. Both exhibit synchronous changes in state deviation and correlation deviation, but their degree of abnormality and duration do not reach the quality anomaly levels corresponding to B04 and B09. Therefore, the system retains them as quality concern states. This result indicates that during packaging quality monitoring, sensor state deviation is used to identify anomalies within a single node, while sensor correlation deviation is used to identify changes in node coordination relationships. The combination of these two methods can distinguish between different quality scenarios, including normal fluctuations, single-node state changes, abnormal correlation relationships, and combined anomalies in both state and correlation.

[0168] Through the above implementation process, the dynamic correlation diagram of packaging quality can adjust the node relationships according to the forming, filling, sealing and conveying processes. The GDN diagram deviation network aggregates the effective adjacent node states under the constraints of the current packaging process and forms a dual deviation representation based on the node prediction state and the node correlation state. The monitoring results and abnormal node relationships in Table 1 can be directly mapped to specific packaging processes and abnormal locations, which can be used for subsequent quality verification, equipment process parameter investigation and packaging bag quality traceability.

[0169] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of feed packaging bag quality based on sensor networks, characterized in that, Includes the following steps: S1 collects multi-source quality monitoring data and packaging equipment operation status data during the feed packaging process, and forms packaging quality time-series data according to packaging bag identification, node identification and collection time. S2, perform time alignment and process division on the packaging quality time series data, and encode the sensor data of each process stage to obtain the packaging quality state sequence; S3, acquire sensor node attributes and historical monitoring data of qualified packaging, establish the quality impact relationship of packaging process, construct a dynamic correlation diagram of packaging quality and continuously update it to obtain correlation change records; S4. Input the packaging quality status sequence and packaging quality dynamic correlation diagram into the GDN graph deviation network, and obtain the quality status prediction result through process constraint graph attention aggregation; S5. Establish a normal packaging correlation benchmark based on the historical monitoring data of qualified packaging, and calculate the dual deviation by combining the quality status prediction results and correlation change records to obtain the packaging bag quality deviation characteristics. S6. Identify abnormal sensor nodes and associated abnormal nodes based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, and determine the type of quality abnormality and the area of ​​abnormality. S7. Determine the quality status of the packaging bag based on the characteristics of the packaging bag quality deviation, and generate the packaging bag quality monitoring results by combining the identified quality anomaly types and the areas affected by the anomalies.

2. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 1, characterized in that, S2 includes: S21, establish a unified time series based on the collection time in the packaging quality time series data, and match the time data of different sensors for the same packaging bag; S22, determine the stages of forming, loading, sealing and conveying processes based on the operating status of packaging equipment and changes in sensor response, and establish the correspondence between process stages and a unified time series; S23, extract sensor data according to the time range corresponding to each process stage to obtain process stage monitoring data under different process stages; S24. The process stage identifier and process stage monitoring data are jointly coded and arranged according to the sequence of processes to obtain the packaging quality status sequence.

3. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 1, characterized in that, S3 includes: S31. Obtain the sensor type and installation location from the node configuration record, combine them with the packaging process stage to form node attribute characteristics, and extract the node temporal change characteristics from the packaging quality status sequence. S32, the node attribute features and node temporal change features are fused to obtain sensor node features, and the correlation degree of node state change is determined based on the node state change; S33, establish the quality impact relationship of the packaging process based on sensor node attributes and process stages, and select node combinations based on the correlation of node state changes to form an initial quality association set; S34. Determine the reliability of the quality association of node combinations based on historical monitoring data of qualified packaging, adjust the association weights in combination with process stages, construct a dynamic association diagram of packaging quality, and generate association change records.

4. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 3, characterized in that, S34 includes: S341, Time alignment and process division are performed on the historical monitoring data of qualified packaging, and the state change relationship of node combinations in the initial quality association set is extracted to form historical node association data; S342, calculate the degree of common change, the degree of consistency of change direction, and the degree of association persistence of node combinations based on historical node association data to obtain historical association stability characteristics; S343, jointly encode the correlation degree of node state changes, the impact of packaging process quality, and historical correlation stability characteristics to obtain reliable quality correlation characteristics; S344: Based on the reliable characteristics of quality association and the current process stage, select reliable quality association relationships, adjust the node association weights, continuously construct a dynamic association diagram of packaging quality, and record association changes.

5. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 4, characterized in that, S344 includes: S3441, Determine the effective sensor nodes based on the current process stage, and extract the node combination containing the effective sensor nodes from the initial quality association set; S3442, based on the reliable characteristics of quality association and the quality impact relationship of packaging process, filter out node combinations that do not match the current process to obtain reliable quality association relationships; S3443, determine the association weight of credible quality association based on the correlation degree of node state changes, historical association stability characteristics and the impact of packaging process quality; S3444 uses effective sensor nodes as graph nodes, reliable quality association relationships as graph edges, and association weights as graph edge attributes to continuously construct a dynamic association graph of packaging quality, and records association changes according to the monitoring time to obtain association change records.

6. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 1, characterized in that, S4 includes: S41, Time alignment and process coding are performed on the historical monitoring data of qualified packaging to form a historical packaging quality status sequence, and a corresponding dynamic correlation graph and training samples are constructed to train the GDN graph deviation network. S42, extract the node time sequence status before the target monitoring time from the packaging quality status sequence, and determine the effective adjacent nodes and association weights based on the packaging quality dynamic association graph; S43, input the temporal state, association weight and current process stage of the target node and its effective adjacent nodes into the attention layer of the process constraint graph to obtain the association quality aggregation feature; S44 integrates the target node's temporal status, associated quality aggregation features, and the current process stage, and obtains the quality status prediction result for the target monitoring time through the trained GDN graph deviation network.

7. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 6, characterized in that, S43 includes: S431, Extract the set of effective adjacent nodes and node association weights corresponding to the target sensor node based on the dynamic correlation diagram of packaging quality; S432, perform feature matching on the temporal states of the target node and each effective neighboring node to obtain attention response features that characterize the degree of correlation between node quality states; S433 integrates the attention response features with the node association weights and the packaging process quality influence relationship established in S3, suppresses the adjacent responses that do not match the current process, and obtains the process constraint attention weights. S434: Based on the process constraint attention weight, the temporal state of each effective adjacent node is weighted and aggregated to obtain the association quality aggregation feature of the target node.

8. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 1, characterized in that, S5 includes: S51, compare the quality status prediction results with the actual monitoring status corresponding to the target monitoring time, and form the sensor status deviation based on the difference in node status prediction. S52, Time alignment and process division of historical monitoring data of qualified packaging, and establish historical correlation diagram using a diagramming method consistent with the current dynamic correlation diagram of packaging quality, to form a normal packaging correlation benchmark; S53, compare the current node association status with the normal packaging association benchmark, and combine the continuous association changes in the association change record to obtain the sensor association deviation; S54, jointly encodes the degree, direction and duration of sensor state deviation and sensor correlation deviation, retains node identifier, correlation relationship of associated nodes and time information, and obtains the packaging bag quality deviation characteristics.

9. The intelligent monitoring method for feed packaging bag quality based on sensor networks according to claim 1, characterized in that, S6 includes: S61, using the trained GDN graph deviation network to predict the verification samples in the historical monitoring data of qualified packaging, determining the normal deviation range of nodes based on the prediction error, and screening abnormal sensor nodes by combining state deviation, correlation deviation and continuous state. S62, read the correlation relationship of abnormal sensor nodes along the dynamic correlation diagram of packaging quality, and identify the peer node with continuous abnormal correlation and the node with synchronous quality deviation change as the associated abnormal node. S63, based on the sensor type, installation location and process stage of the abnormal node, establish the correspondence between the sensor installation location and the effective area of ​​the packaging bag, and determine the abnormal process and abnormal action location; S64. Determine the quality anomaly type based on the quality impact relationship of the packaging process and the combination of abnormal nodes, and obtain the anomaly area by combining the location of the anomaly; no anomaly type or anomaly area is generated when no anomaly is detected.

10. A sensor network-based intelligent monitoring system for the quality of feed packaging bags, characterized in that, The intelligent monitoring method for feed packaging bag quality based on sensor networks according to any one of claims 1 to 9 includes the following modules: The data acquisition module is used to collect multi-source quality monitoring data and packaging equipment operating status data to form packaging quality time-series data; The state sequence construction module is used to perform time alignment and process division on packaging quality time-series data to obtain a packaging quality state sequence. The dynamic correlation graph construction module is used to obtain sensor node attributes and historical monitoring data of qualified packaging, establish the quality impact relationship of packaging process, and construct a dynamic correlation graph of packaging quality and correlation change records. The GDN quality prediction module is used to train the GDN graph deviation network based on historical monitoring data of qualified packaging, and obtain the quality status prediction result through process constraint graph attention aggregation. The quality deviation analysis module is used to establish a normal packaging correlation benchmark, determine the sensor status deviation based on the quality status prediction results, determine the sensor correlation deviation based on the current node correlation status and correlation change records, and fuse them to obtain the packaging bag quality deviation characteristics. The anomaly location module is used to determine the abnormal sensor nodes and associated abnormal nodes based on the characteristics of packaging bag quality deviation and the dynamic correlation diagram of packaging quality, and to determine the quality anomaly type and the area of ​​anomaly when an anomaly exists. The quality result generation module is used to determine the quality status of packaging bags based on the characteristics of packaging bag quality deviations, and generate packaging bag quality monitoring results by combining the identified quality anomaly types and the areas affected by the anomalies.

Images