Honey pomelo processing quality management method and system based on industrial data management

Abstract

The invention discloses a honey pomelo processing quality management method and system based on industrial data management, and the method comprises the following steps: S1, collecting industrial data, and generating a structured data set; s2, constructing a multi-layer weighted graph, defining nodes and initializing feature vectors and edge weights; s3, executing graph neural network message passing and feature aggregation on the graph, and constructing a correlation model; s4, mapping real-time processing data to a correlation model tensor, predicting quality deviation and evaluating a process state; s5, executing adaptive anomaly detection, and generating an abnormal node set and deviation distribution; s6, comparing the abnormal nodes with historical data to form a traceable abnormal record and a quality file; and S7, node features are updated, a graph structure is optimized, a process adjustment strategy is generated, and closed-loop quality management is realized. The honey pomelo processing quality real-time prediction, anomaly identification and closed-loop management are realized through full-process data analysis and adaptive anomaly detection, and the processing precision and traceability are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial data management technology, and in particular to a method and system for quality management of pomelo processing based on industrial data management. Background Technology

[0002] As an important fruit category, pomelo's processing involves multiple stages, including washing, peeling, cutting, grading, and packaging. The processing parameters, equipment status, and quality indicators at each stage directly affect the final product's quality and market competitiveness. Current pomelo processing quality management methods largely rely on manual experience and decentralized testing. Quality data is mostly recorded on paper or stored in separate spreadsheets, lacking unified collection, processing, and analysis of real-time data from each processing stage. This results in significant quality fluctuations during processing and makes it difficult to detect anomalies in a timely manner.

[0003] Some studies have attempted to apply information technology to pomelo processing, such as using sensors to collect equipment operating status or visual inspection systems to inspect appearance quality. However, these methods have significant shortcomings: First, the data collection scope is limited, usually only covering some stages, making it difficult to achieve full-process quality monitoring. Second, the data processing methods are simplistic, mostly relying on static statistics or manual analysis, lacking structured and standardized processing methods, and failing to form a unified dataset for automatic system calculation. Third, there is a lack of effective correlation between processing technology and final quality indicators; existing methods struggle to predict and analyze quality results based on processing parameters, and cannot form a complete quality traceability system. Fourth, anomaly detection relies on experience-based threshold setting, lacking multi-dimensional, dynamically updated adaptive analysis methods, resulting in the inability to identify and handle anomalies in the processing stage in a timely manner.

[0004] Therefore, how to provide a method and system for quality management of pomelo processing based on industrial data management is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for quality management of pomelo processing based on industrial data management. This invention fully utilizes industrial data acquisition, graph neural networks, multidimensional feature mapping, and adaptive anomaly detection technology. It describes in detail the whole-process structured data processing from raw material information, processing parameters, equipment status data to quality inspection data, the construction of a correlation model between processing technology and quality indicators, real-time quality deviation prediction, and processing technology status assessment method. It realizes automatic identification of anomalies in the processing stage and generation of process adjustment strategies, and has the advantages of strong real-time performance, high prediction accuracy, traceability of processing quality, and high efficiency of closed-loop management.

[0006] A method for quality management of pomelo processing based on industrial data management according to an embodiment of the present invention includes the following steps: S1. Collect raw material information, processing parameters, equipment status data and quality inspection data for the entire pomelo processing process, and perform time synchronization and standardization processing on the collected data to generate a structured dataset; S2. Construct the structured dataset into a multi-layer weighted graph structure, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; S3. Perform parallel message passing and multi-round node feature aggregation on the weighted graph structure to generate processing technology feature representation and quality feature representation, and build a correlation model between processing technology and quality indicators; S4. Perform feature mapping on real-time processing data, and perform tensor operations on node feature vectors and associated models to realize quality deviation prediction and processing process status evaluation of each processing stage. S5. Based on node feature vectors, execute an adaptive anomaly detection algorithm to perform multi-dimensional analysis on abnormal nodes and generate a set of abnormal nodes and deviation distribution information in the processing stage. S6. Compare the set of abnormal nodes with historical batch data to create a comparison chart, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality characteristic archives. S7. Based on the abnormal records and node feature updates, execute the graph structure optimization algorithm to generate processing technology adjustment strategies, map the adjustment strategies to processing nodes, and realize closed-loop quality management.

[0007] Optionally, S1 specifically includes: S11. The raw material information includes the source, variety, maturity, weight, sugar content, acidity, and surface defects of the pomelo. The batch identification of the raw material is read by barcode or RFID, and the appearance image is collected by high-definition camera for color, size and defect analysis. At the same time, the harvest date and transportation conditions are recorded. S12. The processing parameters include water temperature, water flow rate, processing equipment speed, cutting blade angle, sorting vibration frequency, and packaging pressure in the cleaning, peeling, cutting, grading, and packaging stages. Real-time control signals and setting parameters are collected by PLC or sensors, and the timestamps are synchronized uniformly. S13. The equipment status data includes the current, voltage, operating speed, load and vibration of each processing equipment, which are collected through industrial control bus or field sensors, and abnormal fluctuations and shutdown records are marked. S14. The quality inspection data includes the detection values ​​output by the online saccharimeter, acidity meter, hardness meter and visual inspection system, and is recorded according to the processing batch and time sequence. S15. The collected industrial data is processed by denoising, outlier removal, and unit standardization to form a structured dataset.

[0008] Optionally, S2 specifically includes: S21. Index the structured dataset by processing batch, generate corresponding processing nodes for each batch's cleaning, peeling, cutting, grading and packaging steps, and assign a unique identifier to each node. S22. For each processing node, extract processing time, equipment speed, load, temperature, humidity, operation sequence and batch processing volume from the structured dataset. Normalize the numerical features to a uniform numerical range, and use one-hot encoding to generate node feature vectors for categorical features, processing equipment type and operator ID. S23. Calculate the edge weights between processing nodes. For nodes with sequential processing, calculate the weights based on the time interval between the completion time of the preceding node and the start time of the following node, the turnover ratio of the same batch, and the sharing ratio of processing equipment. Store the edge weights in the adjacency matrix. S24. Combine the node feature vectors and edge weights to form the initial representation of each node. When constructing a multi-layer graph structure, map the raw material nodes to the first layer, the processing nodes to the second layer, and the equipment nodes to the third layer by layering different processing stages. The node features of each layer are associated with the nodes of the previous layer and the same layer through the adjacency matrix to form a multi-layer weighted graph structure.

[0009] Optionally, S3 specifically includes: S31. Load the initial feature vector of each processing node into the GPU memory, and establish node index and edge index according to the adjacency matrix of the graph. Store the starting node, ending node and edge weight of each edge into a tensor structure. S32. In each round of message passing, the neighbor node feature weighted accumulation operation is performed sequentially for each node, specifically as follows: Read the neighbor node feature vectors from the edge index, perform element-wise multiplication according to the corresponding edge weights, and then sum the neighbor node feature vectors to obtain the weighted neighbor feature vectors. The weighted neighbor feature vector and the current node feature vector are concatenated in the tensor dimension, input into a linear transformation matrix for matrix multiplication mapping, and then processed by a nonlinear activation function to generate the update features for the current round of nodes. S33. After completing one round of updates, the mean and variance of node features are adjusted by batch normalization, and the results are used as the input for the next round of message passing. This process is repeated for multiple rounds of iteration, with all nodes processed in parallel in each round of iteration, to achieve the accumulation of neighbor information and feature aggregation. S34. After multiple rounds of aggregation, the feature vectors of each node are classified according to the node type. Tensor concatenation is used to merge the features of the processing nodes to form a processing technology feature representation, and the features of the quality inspection nodes are merged to form a quality feature representation. S35. Based on historical batch processing data and quality inspection data, perform dot product operation on the processing technology feature representation and the quality feature representation element by element, calculate the mapping weight between node features and quality indicators, and construct a correlation model between processing technology and quality indicators for real-time quality deviation prediction and process status assessment.

[0010] Optionally, the construction process of the association model specifically includes: Align the processing technology feature representation of the processing node with the quality feature representation of the quality inspection node on the same batch dimension, and form a pair input between the feature vector of each processing node and the quality feature vector of the corresponding batch. A weight tensor is assigned to each pair of feature vectors. The weights are calculated based on the position of the processing node in the processing flow, the batch turnover ratio, and the edge weights between the processing node and the quality inspection node. Tensor dot product operations are performed in parallel on the GPU to multiply the processing technology feature representation and the quality feature representation element by element according to weights and accumulate them to generate a preliminary correlation score between each processing node and the quality index. The initial association scores of all nodes are batch normalized, and the normalization results are input into a fully connected neural network. The fully connected neural network is composed of several linear layers, nonlinear activation layers and batch normalization layers stacked together, and outputs the association weight matrix between the processing nodes and the quality indicators. In the output stage, the association weight matrix of each processing node is matched with its corresponding quality index label to form a mapping relationship between processing technology and quality index. The mapping relationship is then stored as a tensor structure that can be computed in real time, thereby completing the construction of the association model.

[0011] Optionally, S4 specifically includes: S41. Convert the real-time collected processing parameters and quality inspection data into feature vectors of real-time nodes according to batch and processing node index, load the feature vectors into GPU memory, and load the weight tensor of the associated model into the GPU. S42. For each processing node, perform element-wise multiplication of the feature vector of the real-time node with the weight tensor of the associated model, and then accumulate the weighted features of the node's neighbors to generate the node state vector. S43. Based on this, perform tensor dot product operation on the mapping matrix between the node state vector and the associated model to calculate the deviation prediction value of the quality index corresponding to each processing node. S44. Perform batch normalization on the deviation prediction values, and stitch together the prediction results of all processing nodes into the quality deviation tensor of the processing batch according to the processing flow sequence. S45. Through parallel computing with GPU, the quality deviation tensor is combined with the node state vector to generate a processing state evaluation result, which is used to determine whether the current process of each processing step deviates from the set target.

[0012] Optionally, S5 specifically includes: S51. Align the real-time acquired node feature vectors with the historical batch feature vectors in the GPU memory according to batch and node index to form a three-dimensional tensor structure, where the dimensions correspond to the node number, feature dimension, and time series, respectively. S52. Perform sliding window statistical operation on each node feature dimension to calculate the mean, variance and skewness within the window, forming a dynamic statistical feature vector. S53. Subtract the mean of the corresponding window from the node feature vector and divide each element by the standard deviation to obtain the standardized deviation tensor. S54. Based on this, an adaptive anomaly detection algorithm is executed. The adaptive anomaly detection algorithm calculates the anomaly score of each node based on the standardized deviation tensor and combines the weighted feature contribution of neighboring nodes to form a comprehensive anomaly score. S55. Apply adaptive threshold determination to the comprehensive anomaly score of each node on the GPU. The threshold is obtained by calculating the percentile of the historical batch anomaly score sequence. The threshold sequence is updated in each round of determination to adapt to changes in processing characteristics. S56. For nodes that are determined to be abnormal, extract their processing parameters, equipment status characteristics, neighbor node characteristics, and corresponding quality index deviations into the abnormal feature matrix. At the same time, statistically analyze the distribution of deviations of each node on each quality index to form a deviation distribution tensor. S57. In parallel processing on the GPU, all abnormal nodes and deviation distribution information are aggregated to generate a set of abnormal nodes and deviation distribution information for the processing stage.

[0013] Optionally, S6 specifically includes: S61. Load the abnormal node set into GPU memory according to the processing batch and node type index, and construct a historical node set according to the same index for the processing node characteristics, processing equipment status and quality index characteristics of each batch in the historical batch data; S62. Perform parallel matching of the abnormal node set and the historical node set on the GPU to construct a comparison graph. For each abnormal node, calculate the Euclidean distance and weighted similarity with the historical node in terms of processing parameters, processing equipment status characteristics and quality index deviations. Store the distance and similarity as edge weights in the comparison graph. S63. In the comparison diagram, each abnormal node is connected to the historical node through the adjacency matrix, forming a node-edge-node structure. S64. Based on this, perform serialization graph analysis, encode the sequence of abnormal nodes on each processing path through depth-first traversal, and concatenate the processing sequence, abnormal type, deviation value and timestamp information to form a sequence vector. S65. Summarize the sequence vectors of all batches to generate traceable quality anomaly records, and package and store the anomaly node characteristics, processing parameters, processing equipment status and quality index deviations of each batch as a processing batch quality feature file.

[0014] A pomelo processing quality management system based on industrial data management according to an embodiment of the present invention includes: The data acquisition module is used to collect raw material information, processing parameters, equipment status data and quality inspection data throughout the entire pomelo processing process, and to perform time synchronization and standardization processing on the collected data to generate a structured dataset. The graph construction module is used to construct a multi-layer weighted graph structure from a structured dataset, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; The feature aggregation module is used to perform parallel message passing and multi-round node feature aggregation on a weighted graph structure to generate processing technology feature representations and quality feature representations, and to build a correlation model between processing technology and quality indicators. The computation and evaluation module is used to perform feature mapping on real-time processing data, perform tensor operations on node feature vectors and associated models, and realize the prediction of quality deviations and the evaluation of processing status in each processing stage. The anomaly detection module is used to execute an adaptive anomaly detection algorithm based on node feature vectors, perform multi-dimensional analysis on abnormal nodes, and generate a set of abnormal nodes and deviation distribution information in the processing stage. The comparative analysis module is used to create a comparison chart between the set of abnormal nodes and historical batch data, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality feature files. The optimization and adjustment module is used to update abnormal records and node characteristics, execute graph structure optimization algorithms to generate processing technology adjustment strategies, and map the adjustment strategies to processing nodes to achieve closed-loop quality management.

[0015] The beneficial effects of this invention are: First, by comprehensively collecting, standardizing, and structurally storing raw material information, processing parameters, equipment status data, and quality inspection data throughout the entire pomelo processing process, this invention achieves unified management and full-process coverage of processing data, providing a complete and reliable data foundation for subsequent analysis and quality assessment.

[0016] Secondly, by using a multi-layer weighted graph structure and graph neural network to perform parallel message passing and multi-round aggregation on node features, this invention can construct a correlation model between processing technology and quality indicators, perform feature mapping and tensor operations on real-time processing data, realize quality deviation prediction and process status evaluation of each processing link, and generate an abnormal node set and deviation distribution information through adaptive anomaly detection, so that anomalies in the processing link can be detected and located in a timely manner.

[0017] Finally, by comparing and analyzing abnormal node information with historical batch data, and by executing graph structure optimization to generate processing technology adjustment strategies, this invention achieves closed-loop optimization and quality traceability management of the processing technology, making the quality control of pomelo processing more precise and efficient, ensuring product quality stability, and significantly improving the intelligence level and management efficiency of the processing process. Attached Figure Description

[0018] 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 The flowchart shows a pomelo processing quality management method based on industrial data management proposed in this invention. Figure 2 This is a flowchart of real-time quality deviation prediction and closed-loop process adjustment for a pomelo processing quality management method based on industrial data management proposed in this invention. Figure 3 This is a module structure diagram of a pomelo processing quality management system based on industrial data management proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1-2 A method for quality management in pomelo processing based on industrial data management includes the following steps: S1. Collect raw material information, processing parameters, equipment status data and quality inspection data for the entire pomelo processing process, and perform time synchronization and standardization processing on the collected data to generate a structured dataset; S2. Construct the structured dataset into a multi-layer weighted graph structure, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; S3. Perform parallel message passing and multi-round node feature aggregation on the weighted graph structure to generate processing technology feature representation and quality feature representation, and build a correlation model between processing technology and quality indicators; S4. Perform feature mapping on real-time processing data, and perform tensor operations on node feature vectors and associated models to realize quality deviation prediction and processing process status evaluation of each processing stage. S5. Based on node feature vectors, execute an adaptive anomaly detection algorithm to perform multi-dimensional analysis on abnormal nodes and generate a set of abnormal nodes and deviation distribution information in the processing stage. S6. Compare the set of abnormal nodes with historical batch data to create a comparison chart, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality characteristic archives. S7. Based on the abnormal records and node feature updates, execute the graph structure optimization algorithm to generate processing technology adjustment strategies, map the adjustment strategies to processing nodes, and realize closed-loop quality management.

[0021] In this embodiment, S1 specifically includes: S11. The raw material information includes the source, variety, maturity, weight, sugar content, acidity, and surface defects of the pomelo. The batch identification of the raw material is read by barcode or RFID, and the appearance image is collected by high-definition camera for color, size and defect analysis. At the same time, the harvest date and transportation conditions are recorded. S12. The processing parameters include water temperature, water flow rate, processing equipment speed, cutting blade angle, sorting vibration frequency, and packaging pressure in the cleaning, peeling, cutting, grading, and packaging stages. Real-time control signals and setting parameters are collected by PLC or sensors, and the timestamps are synchronized uniformly. S13. The equipment status data includes the current, voltage, operating speed, load and vibration of each processing equipment, which are collected through industrial control bus or field sensors, and abnormal fluctuations and shutdown records are marked. S14. The quality inspection data includes the detection values ​​output by the online saccharimeter, acidity meter, hardness meter and visual inspection system, and is recorded according to the processing batch and time sequence. S15. The collected industrial data is processed by denoising, outlier removal, and unit standardization to form a structured dataset.

[0022] In this embodiment, S2 specifically includes: S21. Index the structured dataset by processing batch, generate corresponding processing nodes for each batch's cleaning, peeling, cutting, grading and packaging steps, and assign a unique identifier to each node. S22. For each processing node, extract processing time, equipment speed, load, temperature, humidity, operation sequence and batch processing volume from the structured dataset. Normalize the numerical features to a uniform numerical range, and use one-hot encoding to generate node feature vectors for categorical features, processing equipment type and operator ID. S23. Calculate the edge weights between processing nodes. For nodes with sequential processing, calculate the weights based on the time interval between the completion time of the preceding node and the start time of the following node, the turnover ratio of the same batch, and the sharing ratio of processing equipment. Store the edge weights in the adjacency matrix. S24. Combine the node feature vectors and edge weights to form the initial representation of each node. When constructing a multi-layer graph structure, map the raw material nodes to the first layer, the processing nodes to the second layer, and the equipment nodes to the third layer by layering different processing stages. The node features of each layer are associated with the nodes of the previous layer and the same layer through the adjacency matrix to form a multi-layer weighted graph structure.

[0023] In this embodiment, S3 specifically includes: S31. Load the initial feature vector of each processing node into the GPU memory, and establish node index and edge index according to the adjacency matrix of the graph. Store the starting node, ending node and edge weight of each edge into a tensor structure. S32. In each round of message passing, the neighbor node feature weighted accumulation operation is performed sequentially for each node, specifically as follows: Read the neighbor node feature vectors from the edge index, perform element-wise multiplication according to the corresponding edge weights, and then sum the neighbor node feature vectors to obtain the weighted neighbor feature vectors. The weighted neighbor feature vector and the current node feature vector are concatenated in the tensor dimension, input into a linear transformation matrix for matrix multiplication mapping, and then processed by a nonlinear activation function to generate the update features for the current round of nodes. S33. After completing one round of updates, the mean and variance of node features are adjusted by batch normalization, and the results are used as the input for the next round of message passing. This process is repeated for multiple rounds of iteration, with all nodes processed in parallel in each round of iteration, to achieve the accumulation of neighbor information and feature aggregation. S34. After multiple rounds of aggregation, the feature vectors of each node are classified according to the node type. Tensor concatenation is used to merge the features of the processing nodes to form a processing technology feature representation, and the features of the quality inspection nodes are merged to form a quality feature representation. S35. Based on historical batch processing data and quality inspection data, perform dot product operation on the processing technology feature representation and the quality feature representation element by element, calculate the mapping weight between node features and quality indicators, and construct a correlation model between processing technology and quality indicators for real-time quality deviation prediction and process status assessment.

[0024] In this embodiment, the construction process of the association model specifically includes: Align the processing technology feature representation of the processing node with the quality feature representation of the quality inspection node on the same batch dimension, and form a pair input between the feature vector of each processing node and the quality feature vector of the corresponding batch. A weight tensor is assigned to each pair of feature vectors. The weights are calculated based on the position of the processing node in the processing flow, the batch turnover ratio, and the edge weights between the processing node and the quality inspection node. Tensor dot product operations are performed in parallel on the GPU to multiply the processing technology feature representation and the quality feature representation element by element according to weights and accumulate them to generate a preliminary correlation score between each processing node and the quality index. The initial association scores of all nodes are batch normalized, and the normalization results are input into a fully connected neural network. The fully connected neural network is composed of several linear layers, nonlinear activation layers and batch normalization layers stacked together, and outputs the association weight matrix between the processing nodes and the quality indicators. In the output stage, the association weight matrix of each processing node is matched with its corresponding quality index label to form a mapping relationship between processing technology and quality index. The mapping relationship is then stored as a tensor structure that can be computed in real time, thereby completing the construction of the association model.

[0025] In this embodiment, S4 specifically includes: S41. Convert the real-time collected processing parameters and quality inspection data into feature vectors of real-time nodes according to batch and processing node index, load the feature vectors into GPU memory, and load the weight tensor of the associated model into the GPU. S42. For each processing node, perform element-wise multiplication of the feature vector of the real-time node with the weight tensor of the associated model, and then accumulate the weighted features of the node's neighbors to generate the node state vector. S43. Based on this, perform tensor dot product operation on the mapping matrix between the node state vector and the associated model to calculate the deviation prediction value of the quality index corresponding to each processing node. S44. Perform batch normalization on the deviation prediction values, and stitch together the prediction results of all processing nodes into the quality deviation tensor of the processing batch according to the processing flow sequence. S45. Through GPU parallel computing, the quality deviation tensor is combined with the node state vector to generate a processing technology status evaluation result, which is used to determine whether the current process of each processing stage deviates from the set target. Specifically, this includes: Align the node state vector and the quality deviation tensor on the same tensor dimension, and perform element-wise multiplication of the node state features and deviation prediction values ​​to calculate the weighted process deviation vector. On the GPU, the weighted process deviation vectors of all processing nodes within the same processing batch are summed along the node dimension to obtain the overall deviation tensor of the batch-level processing process. Batch normalization and sliding time window processing are applied to the overall deviation tensor, and the historical deviation trend of each processing stage is combined with the real-time deviation to generate the deviation time series of the processing node. By performing parallel matrix operations, the deviation time series and node feature vectors are input into the multilayer perceptron network to perform nonlinear mapping, and the process status score vector of each processing node is output. The process status score vectors of all nodes are summarized in the order of the processing flow to form a batch-level processing process status evaluation matrix, which is then stored in association with the quality deviation tensor to form the processing process status evaluation result.

[0026] In this embodiment, S5 specifically includes: S51. Align the real-time acquired node feature vectors with the historical batch feature vectors in the GPU memory according to batch and node index to form a three-dimensional tensor structure, where the dimensions correspond to the node number, feature dimension, and time series, respectively. S52. Perform sliding window statistical operation on each node feature dimension to calculate the mean, variance and skewness within the window, forming a dynamic statistical feature vector. S53. Subtract the mean of the corresponding window from the node feature vector and divide each element by the standard deviation to obtain the standardized deviation tensor. S54. Based on this, an adaptive anomaly detection algorithm is executed. The adaptive anomaly detection algorithm calculates the anomaly score of each node based on the standardized deviation tensor, and combines it with the weighted feature contributions of neighboring nodes to form a comprehensive anomaly score, specifically including: The standardized deviation tensor is processed in parallel on the GPU in batches by node. The standardized deviation of each node's feature dimension is weighted and accumulated with the feature vectors of its neighboring nodes according to the edge weights to generate a weighted neighbor deviation tensor. The weighted neighbor bias tensor is concatenated with the node's own standardized bias tensor and input into a multi-layer feedforward neural network in parallel on the GPU for feature mapping. Each layer includes linear transformation, batch normalization and non-linear activation, and outputs a preliminary anomaly score for each node. The initial anomaly score is then dynamically normalized using a sliding window of historical anomaly score sequences. The ratio of deviation to mean is calculated, and historical score trends are integrated according to preset weights to generate a comprehensive anomaly score. S55. Apply adaptive threshold determination to the comprehensive anomaly score of each node on the GPU. The threshold is obtained by calculating the percentile of the historical batch anomaly score sequence. The threshold sequence is updated in each round of determination to adapt to changes in processing characteristics. S56. For nodes that are determined to be abnormal, extract their processing parameters, equipment status characteristics, neighbor node characteristics, and corresponding quality index deviations into the abnormal feature matrix. At the same time, statistically analyze the distribution of deviations of each node on each quality index to form a deviation distribution tensor. S57. In parallel processing on the GPU, all abnormal nodes and deviation distribution information are aggregated to generate a set of abnormal nodes and deviation distribution information for the processing stage.

[0027] In this embodiment, S6 specifically includes: S61. Load the abnormal node set into GPU memory according to the processing batch and node type index, and construct a historical node set according to the same index for the processing node characteristics, processing equipment status and quality index characteristics of each batch in the historical batch data; S62. Perform parallel matching of the abnormal node set and the historical node set on the GPU to construct a comparison graph. For each abnormal node, calculate the Euclidean distance and weighted similarity with the historical node in terms of processing parameters, processing equipment status characteristics and quality index deviations. Store the distance and similarity as edge weights in the comparison graph. S63. In the comparison diagram, each abnormal node is connected to the historical node through the adjacency matrix, forming a node-edge-node structure. S64. Based on this, perform serialization graph analysis, encode the sequence of abnormal nodes on each processing path through depth-first traversal, and concatenate the processing sequence, abnormal type, deviation value and timestamp information to form a sequence vector. S65. Summarize the sequence vectors of all batches to generate traceable quality anomaly records, and package and store the anomaly node characteristics, processing parameters, processing equipment status and quality index deviations of each batch as a processing batch quality feature file.

[0028] refer to Figure 3 A pomelo processing quality management system based on industrial data management includes: The data acquisition module is used to collect raw material information, processing parameters, equipment status data and quality inspection data throughout the entire pomelo processing process, and to perform time synchronization and standardization processing on the collected data to generate a structured dataset. The graph construction module is used to construct a multi-layer weighted graph structure from a structured dataset, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; The feature aggregation module is used to perform parallel message passing and multi-round node feature aggregation on a weighted graph structure to generate processing technology feature representations and quality feature representations, and to build a correlation model between processing technology and quality indicators. The computation and evaluation module is used to perform feature mapping on real-time processing data, perform tensor operations on node feature vectors and associated models, and realize the prediction of quality deviations and the evaluation of processing status in each processing stage. The anomaly detection module is used to execute an adaptive anomaly detection algorithm based on node feature vectors, perform multi-dimensional analysis on abnormal nodes, and generate a set of abnormal nodes and deviation distribution information in the processing stage. The comparative analysis module is used to create a comparison chart between the set of abnormal nodes and historical batch data, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality feature files. The optimization and adjustment module is used to update abnormal records and node characteristics, execute graph structure optimization algorithms to generate processing technology adjustment strategies, and map the adjustment strategies to processing nodes to achieve closed-loop quality management.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the entire process quality management of a pomelo processing workshop. In this scenario, from raw material receiving, cleaning, peeling, cutting, grading to packaging, each step of the pomelo process requires strict control of processing parameters and final product quality. However, traditional methods suffer from data silos, large quality fluctuations, and difficulty in tracking anomalies. This invention comprehensively collects raw material information, processing parameters, equipment status data, and quality inspection data in this scenario, and performs time synchronization, standardization, and structured storage of the data to form a complete dataset for analysis. Raw material information includes the weight, sugar content, acidity, size, and surface defects of the pomelo. Processing parameters for cleaning, peeling, cutting, and grading include water temperature, water flow rate, equipment rotation speed, cutter angle, vibration frequency, and batch throughput. Equipment status data includes current, voltage, load, operating speed, and vibration. Quality inspection data includes sugar content, acidity, hardness, color, and appearance defect scores. Through full-process data collection, this invention solves the problem that traditional methods cannot grasp the quality information of the processing process in real time.

[0030] In this scenario, the present invention constructs a multi-layer weighted graph from a structured dataset, defining raw material nodes, processing nodes, and equipment nodes. Node feature vectors are initialized based on node features and edge weights. Parallel message passing and multi-round node feature aggregation are then performed on a graph neural network to generate processing technology feature representations and quality feature representations, and a correlation model between processing technology and quality indicators is constructed. After real-time processing data is input into the system, tensor operations are performed between the node feature vectors and the correlation model to predict quality deviations and assess process status at each processing stage. Adaptive detection of abnormal nodes is performed, generating a set of abnormal nodes and deviation distribution information. The system compares and analyzes abnormal nodes with historical batch data to form a processing batch quality archive. Graph structure optimization is then executed to generate process adjustment strategies, which are mapped to processing nodes to achieve closed-loop quality management. In actual operation, the system can detect abnormal nodes in real time during processing, such as weight deviations exceeding 5 grams, sugar content deviations exceeding 0.3°Brix, or defect rates exceeding 2%, and automatically label the processing batches.

[0031] During implementation, the system monitored 12 processing lines, processing approximately 20,000 pomelos daily. Data analysis revealed a significant reduction in processing deviations. After 30 consecutive days of operation, the average processing deviation decreased from 5.2% using traditional methods to 1.4%, the sugar content fluctuation range decreased from ±0.6°Brix to ±0.15°Brix, the defect rate decreased from an average of 3.8% to 1.2%, the processing anomaly detection rate increased from approximately 65% ​​to 98%, the response time for processing technology adjustments was shortened from an average of 15 minutes to 3 minutes, and data traceability reached 100%. These data fully validate the significant effectiveness of this invention in real-time quality control, anomaly detection, and closed-loop process optimization.

[0032] To more intuitively demonstrate the implementation effect of the present invention, the processing line operation data is summarized in the following table: Table 1 Comparison of Pomelo Processing Quality and Technology Control Data

[0033] As can be seen from the data in Table 1, the present invention is significantly superior to traditional methods in controlling processing deviations, sugar content fluctuations, and defect rates. At the same time, it has a high anomaly identification rate, rapid response to process adjustments, and complete data traceability, which verifies the feasibility and effectiveness of the present invention in actual processing scenarios.

[0034] The above description is only a preferred embodiment 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 quality management in pomelo processing based on industrial data management, characterized in that, Includes the following steps: S1. Collect raw material information, processing parameters, equipment status data and quality inspection data for the entire pomelo processing process, and perform time synchronization and standardization processing on the collected data to generate a structured dataset; S2. Construct the structured dataset into a multi-layer weighted graph structure, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; S3. Perform parallel message passing and multi-round node feature aggregation on the weighted graph structure to generate processing technology feature representation and quality feature representation, and build a correlation model between processing technology and quality indicators; S4. Perform feature mapping on real-time processing data, and perform tensor operations on node feature vectors and associated models to realize quality deviation prediction and processing process status evaluation of each processing stage. S5. Based on node feature vectors, execute an adaptive anomaly detection algorithm to perform multi-dimensional analysis on abnormal nodes and generate a set of abnormal nodes and deviation distribution information in the processing stage. S6. Compare the set of abnormal nodes with historical batch data to create a comparison chart, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality characteristic archives. S7. Based on the abnormal records and node feature updates, execute the graph structure optimization algorithm to generate processing technology adjustment strategies, map the adjustment strategies to processing nodes, and realize closed-loop quality management.

2. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S1 specifically includes: S11. The raw material information includes the source, variety, maturity, weight, sugar content, acidity, and surface defects of the pomelo. The batch identification of the raw material is read by barcode or RFID, and the appearance image is collected by high-definition camera for color, size and defect analysis. At the same time, the harvest date and transportation conditions are recorded. S12. The processing parameters include water temperature, water flow rate, processing equipment speed, cutting blade angle, sorting vibration frequency, and packaging pressure in the cleaning, peeling, cutting, grading, and packaging stages. Real-time control signals and setting parameters are collected by PLC or sensors, and the timestamps are synchronized uniformly. S13. The equipment status data includes the current, voltage, operating speed, load and vibration of each processing equipment, which are collected through industrial control bus or field sensors, and abnormal fluctuations and shutdown records are marked. S14. The quality inspection data includes the detection values ​​output by the online saccharimeter, acidity meter, hardness meter and visual inspection system, and is recorded according to the processing batch and time sequence. S15. The collected industrial data is processed by denoising, outlier removal, and unit standardization to form a structured dataset.

3. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S2 specifically includes: S21. Index the structured dataset by processing batch, generate corresponding processing nodes for each batch's cleaning, peeling, cutting, grading and packaging steps, and assign a unique identifier to each node. S22. For each processing node, extract processing time, equipment speed, load, temperature, humidity, operation sequence and batch processing volume from the structured dataset. Normalize the numerical features to a uniform numerical range, and use one-hot encoding to generate node feature vectors for categorical features, processing equipment type and operator ID. S23. Calculate the edge weights between processing nodes. For nodes with sequential processing, calculate the weights based on the time interval between the completion time of the preceding node and the start time of the following node, the turnover ratio of the same batch, and the sharing ratio of processing equipment. Store the edge weights in the adjacency matrix. S24. Combine the node feature vectors and edge weights to form the initial representation of each node. When constructing a multi-layer graph structure, map the raw material nodes to the first layer, the processing nodes to the second layer, and the equipment nodes to the third layer by layering different processing stages. The node features of each layer are associated with the nodes of the previous layer and the same layer through the adjacency matrix to form a multi-layer weighted graph structure.

4. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S3 specifically includes: S31. Load the initial feature vector of each processing node into the GPU memory, and establish node index and edge index according to the adjacency matrix of the graph. Store the starting node, ending node and edge weight of each edge into a tensor structure. S32. In each round of message passing, the neighbor node feature weighted accumulation operation is performed sequentially for each node, specifically as follows: Read the neighbor node feature vectors from the edge index, perform element-wise multiplication according to the corresponding edge weights, and then sum the neighbor node feature vectors to obtain the weighted neighbor feature vectors. The weighted neighbor feature vector and the current node feature vector are concatenated in the tensor dimension, input into a linear transformation matrix for matrix multiplication mapping, and then processed by a nonlinear activation function to generate the update features for the current round of nodes. S33. After completing one round of updates, the mean and variance of node features are adjusted by batch normalization, and the results are used as the input for the next round of message passing. This process is repeated for multiple rounds of iteration, with all nodes processed in parallel in each round of iteration, to achieve the accumulation of neighbor information and feature aggregation. S34. After multiple rounds of aggregation, the feature vectors of each node are classified according to the node type. Tensor concatenation is used to merge the features of the processing nodes to form a processing technology feature representation, and the features of the quality inspection nodes are merged to form a quality feature representation. S35. Based on historical batch processing data and quality inspection data, perform dot product operation on the processing technology feature representation and the quality feature representation element by element, calculate the mapping weight between node features and quality indicators, and construct a correlation model between processing technology and quality indicators for real-time quality deviation prediction and process status assessment.

5. The method for quality management of pomelo processing based on industrial data management according to claim 4, characterized in that, The construction process of the association model specifically includes: Align the processing technology feature representation of the processing node with the quality feature representation of the quality inspection node on the same batch dimension, and form a pair input between the feature vector of each processing node and the quality feature vector of the corresponding batch. A weight tensor is assigned to each pair of feature vectors. The weights are calculated based on the position of the processing node in the processing flow, the batch turnover ratio, and the edge weights between the processing node and the quality inspection node. Tensor dot product operations are performed in parallel on the GPU to multiply the processing technology feature representation and the quality feature representation element by element according to weights and accumulate them to generate a preliminary correlation score between each processing node and the quality index. The initial association scores of all nodes are batch normalized, and the normalization results are input into a fully connected neural network. The fully connected neural network is composed of several linear layers, nonlinear activation layers and batch normalization layers stacked together, and outputs the association weight matrix between the processing nodes and the quality indicators. In the output stage, the association weight matrix of each processing node is matched with its corresponding quality index label to form a mapping relationship between processing technology and quality index. The mapping relationship is then stored as a tensor structure that can be computed in real time, thereby completing the construction of the association model.

6. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S4 specifically includes: S41. Convert the real-time collected processing parameters and quality inspection data into feature vectors of real-time nodes according to batch and processing node index, load the feature vectors into GPU memory, and load the weight tensor of the associated model into the GPU. S42. For each processing node, perform element-wise multiplication of the feature vector of the real-time node with the weight tensor of the associated model, and then accumulate the weighted features of the node's neighbors to generate the node state vector. S43. Based on this, perform tensor dot product operation on the mapping matrix between the node state vector and the associated model to calculate the deviation prediction value of the quality index corresponding to each processing node. S44. Perform batch normalization on the deviation prediction values, and stitch together the prediction results of all processing nodes into the quality deviation tensor of the processing batch according to the processing flow sequence. S45. Through parallel computing with GPU, the quality deviation tensor is combined with the node state vector to generate a processing state evaluation result, which is used to determine whether the current process of each processing step deviates from the set target.

7. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S5 specifically includes: S51. Align the real-time acquired node feature vectors with the historical batch feature vectors in the GPU memory according to batch and node index to form a three-dimensional tensor structure, where the dimensions correspond to the node number, feature dimension, and time series, respectively. S52. Perform sliding window statistical operation on each node feature dimension to calculate the mean, variance and skewness within the window, forming a dynamic statistical feature vector. S53. Subtract the mean of the corresponding window from the node feature vector and divide each element by the standard deviation to obtain the standardized deviation tensor. S54. Based on this, an adaptive anomaly detection algorithm is executed. The adaptive anomaly detection algorithm calculates the anomaly score of each node based on the standardized deviation tensor and combines the weighted feature contribution of neighboring nodes to form a comprehensive anomaly score. S55. Apply adaptive threshold determination to the comprehensive anomaly score of each node on the GPU. The threshold is obtained by calculating the percentile of the historical batch anomaly score sequence. The threshold sequence is updated in each round of determination to adapt to changes in processing characteristics. S56. For nodes that are determined to be abnormal, extract their processing parameters, equipment status characteristics, neighbor node characteristics, and corresponding quality index deviations into the abnormal feature matrix. At the same time, statistically analyze the distribution of deviations of each node on each quality index to form a deviation distribution tensor. S57. In parallel processing on the GPU, all abnormal nodes and deviation distribution information are aggregated to generate a set of abnormal nodes and deviation distribution information for the processing stage.

8. The method for quality management of pomelo processing based on industrial data management according to claim 1, characterized in that, S6 specifically includes: S61. Load the abnormal node set into GPU memory according to the processing batch and node type index, and construct a historical node set according to the same index for the processing node characteristics, processing equipment status and quality index characteristics of each batch in the historical batch data; S62. Perform parallel matching of the abnormal node set and the historical node set on the GPU to construct a comparison graph. For each abnormal node, calculate the Euclidean distance and weighted similarity with the historical node in terms of processing parameters, processing equipment status characteristics and quality index deviations. Store the distance and similarity as edge weights in the comparison graph. S63. In the comparison diagram, each abnormal node is connected to the historical node through the adjacency matrix, forming a node-edge-node structure. S64. Based on this, perform serialization graph analysis, encode the sequence of abnormal nodes on each processing path through depth-first traversal, and concatenate the processing sequence, abnormal type, deviation value and timestamp information to form a sequence vector. S65. Summarize the sequence vectors of all batches to generate traceable quality anomaly records, and package and store the anomaly node characteristics, processing parameters, processing equipment status and quality index deviations of each batch as a processing batch quality feature file.

9. A pomelo processing quality management system based on industrial data management, comprising the pomelo processing quality management method based on industrial data management as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect raw material information, processing parameters, equipment status data and quality inspection data throughout the entire pomelo processing process, and to perform time synchronization and standardization processing on the collected data to generate a structured dataset. The graph construction module is used to construct a multi-layer weighted graph structure from a structured dataset, define raw material nodes, processing nodes and equipment nodes, and initialize node feature vectors based on node features and edge weights; The feature aggregation module is used to perform parallel message passing and multi-round node feature aggregation on a weighted graph structure to generate processing technology feature representations and quality feature representations, and to build a correlation model between processing technology and quality indicators. The computation and evaluation module is used to perform feature mapping on real-time processing data, perform tensor operations on node feature vectors and associated models, and realize the prediction of quality deviations and the evaluation of processing status in each processing stage. The anomaly detection module is used to execute an adaptive anomaly detection algorithm based on node feature vectors, perform multi-dimensional analysis on abnormal nodes, and generate a set of abnormal nodes and deviation distribution information in the processing stage. The comparative analysis module is used to create a comparison chart between the set of abnormal nodes and historical batch data, perform serialization chart analysis, and generate traceable quality anomaly records and processing batch quality feature files. The optimization and adjustment module is used to update abnormal records and node characteristics, execute graph structure optimization algorithms to generate processing technology adjustment strategies, and map the adjustment strategies to processing nodes to achieve closed-loop quality management.

Images