Batch consistency control method and system based on finished product quality traceability

Abstract

The application relates to the technical field of intelligent manufacturing and industrial big data processing, and discloses a batch consistency control method and system based on finished product quality tracing, wherein the method comprises the following steps: acquiring batch production tracing data and constructing a dynamic multi-relation propagation graph, performing node space-time feature initialization, abnormality propagation simulation, finished product defect reverse tracing and control instruction generation; compared with the prior art which relies on manual experience or static rules for post-tracing, especially under the production conditions of multi-source heterogeneous data interlacing and complex abnormality propagation path, the technical problem that accurate positioning of abnormality sources and timely linkage control cannot be realized is solved; since the production object correlation is represented by a dynamic multi-relation graph, and the multi-relation graph attention propagation and reverse tracing are combined, automatic positioning from a finished product defect to a process, equipment or material batch is realized, and the accuracy and response speed of batch consistency control are improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial big data processing technology, and in particular to a batch consistency control method and system based on finished product quality traceability. Background Technology

[0002] In batch production processes in manufacturing, the quality of finished products is often determined by a combination of factors, including materials, processes, and equipment, and their interrelationships. When quality anomalies occur, it is crucial to quickly and accurately trace the root cause of the anomaly in order to implement timely control measures and ensure overall batch consistency. However, currently widely used batch quality traceability methods are mostly based on manual experience or simple statistical rules, performing isolated analysis of detection data from a single stage, and lacking the ability to uniformly model multiple types of interrelationships in the production process (such as material flow, process dependencies, and equipment load).

[0003] Furthermore, production status changes dynamically over time, and anomalies often propagate along multiple relational paths. Existing methods struggle to characterize this spatiotemporally coupled propagation process, leading to vague or even erroneous root cause localization and hindering effective batch consistency control in complex manufacturing scenarios. Particularly under production conditions involving multiple product varieties, varying batch sizes, and overlapping processes, the lack of automated traceability and control loop mechanisms makes batch quality control highly dependent on post-production manual checks, which are time-consuming, slow to respond, and difficult to guarantee the overall quality stability of the final product batch. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a batch consistency control method based on finished product quality traceability. This method aims to solve the technical problem that existing technologies rely on manual experience or static rules for post-event traceability, especially in production conditions where multi-source heterogeneous data is intertwined and anomaly propagation paths are complex, making it impossible to accurately locate the root cause of anomalies and implement timely linkage control.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a batch consistency control method based on finished product quality traceability.

[0006] The batch consistency control method based on finished product quality traceability includes:

[0007] Step S10: Obtain batch production traceability data, and based on the batch production traceability data, perform the production object relational abstraction task using a dynamic multi-relationship graph construction method to output a dynamic multi-relationship propagation graph;

[0008] Step S20: Based on the dynamic multi-relation propagation graph, perform the node initial representation generation task using a spatiotemporal coding method with relation type constraints, and output the initial spatiotemporal feature embedding;

[0009] Step S30: Based on the initial spatiotemporal feature embedding, a multi-relation graph attention network is used to perform anomaly propagation representation update task, and the final feature representation is output;

[0010] Step S40: Based on the final feature representation, perform the anomaly root cause localization task using the defect node-guided reverse attention tracing method, and output the anomaly root cause probability score;

[0011] Step S50: Generate batch consistency control instructions based on the abnormal root cause probability scores, and output the control instructions.

[0012] Preferably, step S10, which involves acquiring batch production traceability data, performing a production object relational abstraction task based on the batch production traceability data using a dynamic multi-relationship graph construction method, and outputting a dynamic multi-relationship propagation graph, specifically includes:

[0013] Step S101: Obtain entity records from the batch production traceability data, divide the entity records into material batch nodes, process nodes, equipment nodes, and finished product quality inspection nodes, and output the node set;

[0014] Step S102: Based on the node set, read the relationship records in the batch production traceability data, divide the relationship records into material flow relationship, process dependency relationship, equipment bearing relationship and quality attribution relationship, and output the relationship type set;

[0015] Step S103: Generate an edge set based on the node set and the relation type set, and write the static attribute features and dynamic state features of each node into the node feature matrix, and output the dynamic multi-relationship propagation graph; wherein, the dynamic multi-relationship propagation graph includes the node set, the edge set, the relation type set and the node feature matrix.

[0016] Preferably, step S20, which involves performing the node initial representation generation task based on the spatiotemporal coding method constrained by relation type in the dynamic multi-relation propagation graph and outputting the initial spatiotemporal feature embedding, specifically includes:

[0017] Step S201: Read the node feature matrix from the dynamic multi-relationship propagation graph, extract static attribute features and dynamic state features according to the node identifier, and output the node initial feature vector;

[0018] Step S202: Generate a periodic time encoding vector based on the timestamp corresponding to the initial feature vector of the node, and assign a relation-specific transformation matrix to each relation type according to the relation type set, and output the relation constraint time feature vector;

[0019] Step S203: Based on the relational constraint time feature vector, a multilayer perceptron is used to perform a feature mapping task and output the initial spatiotemporal feature embedding.

[0020] Preferably, step S202, which involves generating a periodic time encoding vector based on the timestamp corresponding to the initial feature vector of the node, assigning a relation-specific transformation matrix to each relation type according to the relation type set, and outputting the relation constraint time feature vector, specifically includes:

[0021] Step S2021: Normalize the timestamp to obtain normalized time parameters;

[0022] Step S2022: Write the normalized time parameter into the odd-dimensional periodic component and the even-dimensional periodic component respectively to generate the periodic time encoding vector;

[0023] Step S2023: Bind the node initial feature vector, the periodic time encoding vector, and the relation-specific transformation matrix accordingly to generate the relation constraint time feature vector.

[0024] Preferably, step S30, which involves using a multi-relationship graph attention network to perform anomaly propagation representation update based on the initial spatiotemporal feature embedding and outputting the final feature representation, specifically includes:

[0025] Step S301: Based on the initial spatiotemporal feature embedding, retrieve the neighbor nodes of the target node according to different relationship types in the dynamic multi-relationship propagation graph, and output the relationship neighbor set;

[0026] Step S302: Based on the set of relational neighbors, calculate the association strength between the target node and its neighboring nodes using a relation-specific transformation matrix and a learnable attention vector, and output the relational attention coefficient;

[0027] Step S303: The initial spatiotemporal feature embeddings of neighboring nodes are weighted and aggregated according to the relational attention coefficients, and iteratively updated by a nonlinear activation function to output the final feature representation of each node.

[0028] Preferably, step S40, which involves performing the anomaly root cause localization task based on the final feature representation using a defect node-guided reverse attention tracing method and outputting anomaly root cause probability scores, specifically includes:

[0029] Step S401: Based on the final feature representation, determine the defective nodes in the finished product quality inspection nodes, and extract the reverse reachable subgraph corresponding to the defective node from the dynamic multi-relationship propagation graph, and output the reverse traceability candidate node set;

[0030] Step S402: Based on the final feature representation of each candidate node in the reverse traceability candidate node set, calculate the self-abnormal evidence score and reverse attention weight of each candidate node relative to the defective node, and output the root evidence fusion parameters.

[0031] Step S403: Calculate the anomaly root cause probability score for each candidate node based on the root cause evidence fusion parameters. The calculation formula is as follows:

[0032] ;

[0033] in, Candidate nodes The probability score of the root cause of the anomaly For normalization function, The evidence fusion coefficient, Candidate nodes The score of its own abnormal evidence. For a set of relation types, Candidate nodes In relation types The set of downstream neighbor nodes, For downstream nodes Backtracking to candidate nodes Reverse attention weights, For downstream nodes The probability score of the root cause of the anomaly.

[0034] Preferably, step S50, which involves generating batch consistency control instructions based on the anomaly root cause probability scores and then outputting the control instructions, specifically includes:

[0035] Step S501: Generate an anomaly root cause ranking list based on the anomaly root cause probability score of each node, retain the correspondence between node identifier, node type and anomaly root cause probability score, and output the anomaly root cause ranking result.

[0036] Step S502: Based on a preset probability threshold, filter the ranking results of the abnormal root causes, determine the nodes with abnormal root cause probability scores greater than the preset probability threshold as key abnormal root cause nodes, and output the set of key abnormal root cause nodes.

[0037] Step S503: Generate corresponding control instructions based on the node type of each key anomaly root source node in the set of key anomaly root source nodes; when the node type is an equipment node, generate equipment early warning and maintenance suggestion instructions; when the node type is a material batch node, generate material isolation and supply chain traceability instructions; when the node type is a process node, generate process parameter adjustment instructions; the control instructions include target node identifier, control action type, execution priority, and control parameters.

[0038] This invention also provides a batch consistency control system based on finished product quality traceability, comprising:

[0039] The dynamic graph construction module is used to acquire batch production traceability data, and based on the batch production traceability data, it performs the task of abstracting the relational relationships of production objects using a dynamic multi-relationship graph construction method, and outputs a dynamic multi-relationship propagation graph.

[0040] The spatiotemporal coding module is used to perform the node initial representation generation task based on the spatiotemporal coding method constrained by relation type based on the dynamic multi-relation propagation graph, and output the initial spatiotemporal feature embedding.

[0041] The propagation simulation module is used to perform anomaly propagation representation update task based on the initial spatiotemporal feature embedding using a multi-relation graph attention network, and output the final feature representation;

[0042] The root cause tracing module is used to perform the abnormal root cause localization task based on the final feature representation using a defect node-guided reverse attention tracing method, and output the abnormal root cause probability score.

[0043] The control execution module is used to generate batch consistency control instructions based on the probability score of the abnormal root cause and output the control instructions.

[0044] The present invention also provides a batch consistency control device based on finished product quality traceability. The batch consistency control device based on finished product quality traceability includes: a memory, a processor, and a batch consistency control program based on finished product quality traceability stored in the memory and executable on the processor. When the batch consistency control program based on finished product quality traceability is executed by the processor, it implements the above-described method.

[0045] The present invention also provides a computer program product, the computer program product including a batch consistency control program based on finished product quality traceability, the batch consistency control program based on finished product quality traceability implementing the above method when executed by a processor.

[0046] The beneficial effects of this invention are as follows:

[0047] 1. This invention uses a dynamic multi-relationship graph construction method to abstract production objects such as material batches, processes, equipment, and finished product testing, as well as their various types of relationships, into a multi-relationship graph. This fully preserves the structural information and interactive semantics of the production process, providing rich relational basic data for anomaly tracing and avoiding root cause misjudgments caused by information loss.

[0048] 2. This invention employs spatiotemporal coding constrained by relation type and a multi-relation graph attention network, which can generate initial embeddings of nodes based on different relation types and time period characteristics, and simulate the differentiated propagation process of anomalies along different relation paths. This ensures that the final representation of each node incorporates the anomaly information of its multi-relation neighbors, significantly improving the accuracy of anomaly propagation representation and adaptability to complex dynamic environments.

[0049] 3. The defect node-guided reverse attention tracing mechanism proposed in this invention starts with the finished product defect node and calculates the abnormal root cause probability score of each candidate node in reverse along the relation edge, realizing causal inference from phenomenon to root cause. Compared with traditional rule methods, the accuracy and reliability of root cause location are greatly improved.

[0050] 4. This invention automatically generates control instructions for different node types based on the probability score of abnormal root causes, directly converting the tracing results into executable control actions, forming a closed loop of "tracing-location-control", which effectively reduces manual intervention and improves the real-time performance and automation level of batch consistency control. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating a first embodiment of a batch consistency control method based on finished product quality traceability according to the present invention.

[0052] Figure 2 This is a simulation diagram of anomaly propagation in a multi-relationship graph attention network for a batch consistency control method based on finished product quality traceability according to the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1: As Figure 1 The diagram shown is a flowchart of a first embodiment of a batch consistency control method based on finished product quality traceability according to the present invention.

[0056] In the first embodiment, the batch consistency control method based on finished product quality traceability includes:

[0057] Step S10: Obtain batch production traceability data, and based on the batch production traceability data, perform the production object relational abstraction task using a dynamic multi-relationship graph construction method to output a dynamic multi-relationship propagation graph;

[0058] The batch production traceability data acquired in step S10 includes, but is not limited to, records generated by the Manufacturing Execution System (MES), Quality Management System (QMS), and equipment data acquisition system. Specific entity records cover material batch numbers, process stations, equipment numbers, and the results of each finished product inspection item. The dynamic multi-relationship graph construction method abstracts these entities into material batch nodes, process nodes, equipment nodes, and finished product quality inspection nodes. Based on the relationship records, four types of relationships are defined: material flow relationships, process dependency relationships, equipment carrying relationships, and quality attribution relationships, generating corresponding directed edges. The node feature matrix then writes the static attributes (such as equipment model and material specifications) and dynamic states (such as processing parameters and runtime) of each node, forming a unified dynamic multi-relationship propagation graph, which serves as the structured input for subsequent spatiotemporal coding and propagation simulation.

[0059] By constructing discrete batch traceability data into a graph structure containing heterogeneous nodes and multiple types of relationships, this step can organize the originally isolated detection records and process data into a highly interconnected network. This not only explicitly expresses complex production logic such as "who passed through whom," "who processed it," and "final ownership," but also provides directly computable graph topology and feature information for subsequent graph learning models. This is the data foundation for the accurate traceability achieved by this invention.

[0060] Traditional quality traceability methods typically use batch numbers or product codes for linear indexing, which can only retrieve a few directly related records and cannot express multiple cross-dependencies between processes, equipment, and materials. This step overcomes the shortcomings of traditional methods, such as flattened relationship representation and severe information loss, by constructing a dynamic multi-relationship graph. This allows the abnormal propagation path of any production object to be fully recorded and traversed, thus supporting deeper causal reasoning.

[0061] For example, in a certain batch of circuit board production, the same batch of materials first undergoes the SMT placement process and then enters the reflow soldering process. The equipment used includes a pick-and-place machine and a reflow oven, and the finished product inspection node records solder joint defects. Through this step, a graph containing material nodes, two process nodes, two equipment nodes, and a finished product inspection node can be constructed, and material flow edges, process dependency edges, equipment load edges, and quality attribution edges can be automatically generated. Pick-and-place machine parameters and real-time temperature are written into the node characteristics, preserving complete information for subsequent anomaly propagation analysis.

[0062] Step S20: Based on the dynamic multi-relation propagation graph, perform the node initial representation generation task using a spatiotemporal coding method with relation type constraints, and output the initial spatiotemporal feature embedding;

[0063] In step S20, the node feature matrix is ​​first read from the dynamic multi-relationship propagation graph. The static attribute feature vector and dynamic state feature vector, already written, are extracted according to the node identifier and concatenated to obtain the initial node feature vector. Then, a periodic time encoding vector is generated using the timestamp corresponding to the node. Specifically, sine and cosine functions are used to extend the normalized time, forming odd-dimensional and even-dimensional periodic components to capture the periodic patterns of the production process. Simultaneously, independent relationship-specific transformation matrices are assigned to each relationship type (material flow, process dependence, equipment load, and quality attribution) based on the relationship type set. The initial node feature vector, time encoding vector, and corresponding relationship matrix are bound to obtain the relationship constraint time feature vector. Finally, feature mapping is performed through a multilayer perceptron (MLP), outputting a unified initial spatiotemporal feature embedding for use by the multi-relationship graph attention network.

[0064] After this step, each node obtains an initial representation that integrates its static attributes, dynamic state, and time period characteristics under each relationship. This enables subsequent anomaly propagation simulations to distinguish the differences of the same node under different relationship scenarios, thereby improving the ability to characterize the temporal evolution patterns in the production process.

[0065] Traditional methods typically concatenate feature vectors with time information or completely ignore time modalities, failing to express the different processing states of the same equipment at different time periods, and also failing to distinguish the direction and intensity of time influence under different relational paths. This step, through spatiotemporal encoding constrained by relation type, ensures that time information is injected strictly according to relational semantics, which, compared to traditional approaches, better reflects the dynamics of the actual manufacturing process.

[0066] For example, a certain heating furnace equipment node is divided into two timestamps in a batch of production: a preheating section and a constant temperature section. Through this step, the equipment will generate a spatiotemporal embedding containing different cycle modes of preheating and constant temperature for the processed material node under the "equipment load" relationship. The subsequent model can use this to distinguish between defects caused by insufficient preheating and defects caused by constant temperature deviation, avoiding confusion.

[0067] Step S30: Based on the initial spatiotemporal feature embedding, a multi-relation graph attention network is used to perform anomaly propagation representation update task, and the final feature representation is output;

[0068] In step S30, during the anomaly propagation representation update task, based on the initial spatiotemporal feature embeddings, for each target node in the dynamic multi-relationship propagation graph, its neighboring nodes are retrieved according to four relationship types: material flow, process dependence, equipment load, and quality attribution, forming different relational neighbor sets. Then, for each relationship type, a dedicated transformation matrix is ​​used to linearly map the initial spatiotemporal embeddings of the neighboring nodes, and after concatenation with the target node embeddings, a normalized relationship attention coefficient is calculated using a learnable attention vector. This coefficient represents the strength of anomaly propagation from neighboring nodes to the target node under that relationship. All neighbor embeddings are weighted and summed according to the attention coefficients, and then iteratively updated using a nonlinear activation function (such as ELU) and residual connections, ultimately obtaining the final feature representation of each node that integrates anomaly information from multiple relational neighbors.

[0069] Through iterative updates of the multi-relationship graph attention network, this step ensures that the final representation of any node is no longer an isolated local feature, but rather incorporates the anomalous signals transmitted by its multi-hop neighbors under specific relationships. This can simulate the final distribution of anomalies "spreading" along the production network, providing sufficient contextual information for tracing the source of anomalies.

[0070] Existing graph neural networks often employ the homogeneous graph assumption, ignoring edge type differences, which leads to the averaged propagation of anomalies and makes it difficult to identify critical propagation paths. The multi-relationship graph attention mechanism in this step assigns independent attention parameters and transformation matrices to different relationships, highlighting strong anomaly propagation effects in relationships such as "device bearing," and suppressing noise from irrelevant relationships. This outperforms traditional graph algorithms in locating common-cause failures and serial link anomalies.

[0071] For example, suppose a batch of finished products is detected as defective due to dimensional inconsistencies, involving associated equipment A and B, and upstream material batch C. Given the equipment load relationship, equipment A has a much higher attention coefficient than equipment B, and the material flow relationship shows that batch C originates from equipment A. After this update step, the final characteristics of the equipment A node and the material batch C node will carry strong anomaly signals, providing clear guidance for subsequent root cause localization.

[0072] Step S40: Based on the final feature representation, perform the anomaly root cause localization task using the defect node-guided reverse attention tracing method, and output the anomaly root cause probability score;

[0073] Step S40, when performing the anomaly root cause localization task, firstly determines which nodes are defective based on the finished product quality inspection nodes in the final feature representation. It then extracts a reachable subgraph from the dynamic multi-relationship propagation graph, extending backward along directed edges from the defective nodes, resulting in a set of candidate nodes for reverse tracing containing all possible upstream root causes. For each candidate node, an anomaly evidence score is calculated. This score, obtained from the node's final features using a specialized classifier, represents its posterior probability as a root cause. Simultaneously, learnable weights under each relation type are used to calculate the reverse attention weights tracing back from downstream defective nodes to candidate nodes. Finally, the anomaly root cause probability scores of the candidate nodes are iteratively calculated, fusing their own evidence with downstream backpropagation evidence until convergence, outputting the root cause probability of all candidate nodes.

[0074] This step transforms the location of the root cause of an anomaly into a causal inference problem based on graph backpropagation by using defect node-driven reverse tracing. By using observational evidence from downstream defect nodes to propagate upwards layer by layer, the most likely starting cause node can be inferred from the phenomenon of the final defect, thus achieving causal tracing.

[0075] Traditional tracing methods typically employ simple threshold judgments or one-way association rules, which cannot distinguish the confidence levels of multiple candidate root causes and are prone to misjudgment when confounding variables are present. This invention utilizes a reverse attention mechanism capable of reasoning about the allocation of anomaly responsibility within a graph structure. By combining its own anomaly evidence with the reverse weights of downstream nodes, it quantifies and ranks the root cause probabilities, significantly improving the accuracy and interpretability of the localization compared to traditional methods.

[0076] For example, when multiple batches of finished products are found to have insufficient strength defects during finished product inspection, this step calculates the probability score of the shared heat treatment furnace node to be 0.92, while that of a specific mold node is only 0.11. Even though both are related to the finished product, the system can clearly point out that the heat treatment furnace is the root cause, rather than misjudging it as a mold problem, thereby guiding maintenance personnel to prioritize the repair of the heat treatment furnace.

[0077] Step S50: Generate batch consistency control instructions based on the abnormal root cause probability scores, and output the control instructions.

[0078] In step S50, the probability scores of the root causes of anomalies at each node output in step S40 are first sorted from high to low to generate a ranking list containing node identifiers, node types, and scores. Then, the ranking results are filtered using a preset probability threshold (e.g., 0.7), retaining only nodes with scores above the threshold as the set of key root cause nodes. Next, instruction generation logic is triggered according to node type: when it is an equipment node, an equipment warning and maintenance suggestion instruction is generated, carrying the equipment number, preset maintenance parameters, and emergency repair procedures; when it is a material batch node, a material isolation and supply chain traceability instruction is generated, instructing the isolation of the batch of materials and initiating a traceability request to the upstream supplier; when it is a process node, a process parameter adjustment instruction is generated, providing a suggested correction range for the process parameters. All control instructions include the target node identifier, control action type, execution priority, and control parameters, and can be directly issued to the MES or equipment control system.

[0079] This step directly transforms the probabilistic root cause inference results generated by the graph learning model into on-site executable automated operation instructions, connecting the last link from quality data analysis to production control, realizing closed-loop management of batch consistency, minimizing the time delay from manual analysis to decision-making, and preventing further expansion of defects.

[0080] Existing technologies often require personnel confirmation and manual work order creation after problem location, resulting in slow response times and a high risk of errors. This step achieves automatic mapping and hierarchical generation of root causes to control instructions, and can adaptively select the optimal control strategy based on the root cause type. This is superior to traditional methods based on fixed rules or entirely manual processing, and can quickly correct deviations with minimal or no production interruption, ensuring batch consistency.

[0081] For example, if the tracing results show that the root cause is an abnormality in the temperature control unit of a certain injection molding equipment (probability 0.95), the system will immediately generate an equipment warning command, requiring the operator to perform an inspection within 10 minutes and automatically retrieve historical temperature curves as a reference. At the same time, the next batch of tasks for that equipment will be temporarily suspended, and can only continue after maintenance confirmation, to prevent subsequent batch quality accidents.

[0082] Example 2: Furthermore, the present invention provides a batch consistency control system based on finished product quality traceability, which employs a batch consistency control method based on finished product quality traceability in the above embodiments, and can solve a technical problem of batch consistency control based on finished product quality traceability. The beneficial effects of the batch consistency control system based on finished product quality traceability provided by the present invention are the same as those of the batch consistency control method based on finished product quality traceability provided in the above embodiments, and other technical features in the batch consistency control system based on finished product quality traceability are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0083] Example 3: This invention provides a batch consistency control device based on finished product quality traceability. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which are then executed to enable the at least one processor to perform the batch consistency control method based on finished product quality traceability described in Example 1. The batch consistency control device based on finished product quality traceability in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This batch consistency control device based on finished product quality traceability is merely an example and should not limit the functionality or scope of the invention. The device may also include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes based on a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of a batch consistency control device based on finished product quality traceability. The processing unit, read-only memory, and random access memory are interconnected via a bus. The I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including, for example, magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the batch consistency control device based on finished product quality traceability to communicate wirelessly or wiredly with other devices to exchange data. While a batch consistency control device based on finished product quality traceability with various systems has been described, it should be understood that it is not required to implement or possess all the systems described. Alternatively, more or fewer systems can be implemented.

[0084] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the batch consistency control method based on finished product quality traceability as described above. The computer program product provided by this invention can solve a technical problem of batch consistency control based on finished product quality traceability. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as the beneficial effects of the batch consistency control method based on finished product quality traceability provided in the above embodiments, and will not be repeated here.

[0085] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0086] For example, such as Figure 2 The diagram shown is a simulation of anomaly propagation in a multi-relationship graph attention network (MR-GATN) for batch consistency control based on finished product quality traceability, illustrating the technical effect of the MR-GATN anomaly propagation simulation in step S30. The heatmap in the upper left corner displays the MR-GATN coefficients. Darker colors indicate higher attention weights, reflecting the differences in the influence of neighboring nodes on the central node under different relationship types (flow, dependency, and carrying). The 3D surface plot in the upper right corner illustrates the non-linear activation function. The output surface simulates the aggregation and activation process, and its shape reflects the influence of nonlinear transformations on the features. The line graph in the lower left corner shows the change in the feature representation norm with network layer iterations; the norm gradually increases and stabilizes, indicating that the network is effectively learning the propagation pattern. The heatmap in the lower right corner shows the final feature representation. The matrix uses SymLogNorm color coding to handle positive and negative values, highlighting the differences between abnormal and normal features. Compared to traditional methods that treat production processes as isolated units or use static parameter models, this step simulates the propagation effect of anomalies along the production chain through a multi-relationship graph attention network. This dynamically captures the dependencies between nodes. The three colored dots in the graph mark the multi-relationship aggregation mechanism, improving the accuracy and adaptability of anomaly propagation modeling.

[0087] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A batch consistency control method based on finished product quality traceability, characterized in that, The method includes: Step S10: Obtain batch production traceability data, and based on the batch production traceability data, perform the production object relational abstraction task using a dynamic multi-relationship graph construction method to output a dynamic multi-relationship propagation graph; Step S20: Based on the dynamic multi-relation propagation graph, perform the node initial representation generation task using a spatiotemporal coding method with relation type constraints, and output the initial spatiotemporal feature embedding; Step S30: Based on the initial spatiotemporal feature embedding, a multi-relation graph attention network is used to perform anomaly propagation representation update task, and the final feature representation is output; Step S40: Based on the final feature representation, perform the anomaly root cause localization task using the defect node-guided reverse attention tracing method, and output the anomaly root cause probability score; Step S50: Generate batch consistency control instructions based on the abnormal root cause probability scores, and output the control instructions.

2. The batch consistency control method based on finished product quality traceability as described in claim 1, characterized in that, Step S10 involves acquiring batch production traceability data, performing a production object relational abstraction task based on the batch production traceability data using a dynamic multi-relationship graph construction method, and outputting a dynamic multi-relationship propagation graph. Specifically, this includes: Step S101: Obtain entity records from the batch production traceability data, divide the entity records into material batch nodes, process nodes, equipment nodes, and finished product quality inspection nodes, and output the node set; Step S102: Based on the node set, read the relationship records in the batch production traceability data, divide the relationship records into material flow relationship, process dependency relationship, equipment bearing relationship and quality attribution relationship, and output the relationship type set; Step S103: Generate an edge set based on the node set and the relation type set, and write the static attribute features and dynamic state features of each node into the node feature matrix, and output the dynamic multi-relationship propagation graph; wherein, the dynamic multi-relationship propagation graph includes the node set, the edge set, the relation type set and the node feature matrix.

3. The batch consistency control method based on finished product quality traceability as described in claim 1, characterized in that, In step S20, the step of performing the node initial representation generation task based on the spatiotemporal coding method constrained by relation type based on the dynamic multi-relation propagation graph, and outputting the initial spatiotemporal feature embedding, specifically includes: Step S201: Read the node feature matrix from the dynamic multi-relationship propagation graph, extract static attribute features and dynamic state features according to the node identifier, and output the node initial feature vector; Step S202: Generate a periodic time encoding vector based on the timestamp corresponding to the initial feature vector of the node, and assign a relation-specific transformation matrix to each relation type according to the relation type set, and output the relation constraint time feature vector; Step S203: Based on the relational constraint time feature vector, a multilayer perceptron is used to perform a feature mapping task and output the initial spatiotemporal feature embedding.

4. The batch consistency control method based on finished product quality traceability as described in claim 3, characterized in that, Step S202, which involves generating a periodic time encoding vector based on the timestamp corresponding to the initial feature vector of the node, assigning a relation-specific transformation matrix to each relation type according to the relation type set, and outputting the relation constraint time feature vector, specifically includes: Step S2021: Normalize the timestamp to obtain normalized time parameters; Step S2022: Write the normalized time parameter into the odd-dimensional periodic component and the even-dimensional periodic component respectively to generate the periodic time encoding vector; Step S2023: Bind the node initial feature vector, the periodic time encoding vector, and the relation-specific transformation matrix accordingly to generate the relation constraint time feature vector.

5. The batch consistency control method based on finished product quality traceability as described in claim 1, characterized in that, Step S30, which involves using a multi-relationship graph attention network to perform anomaly propagation representation update based on the initial spatiotemporal feature embedding and outputting the final feature representation, specifically includes: Step S301: Based on the initial spatiotemporal feature embedding, retrieve the neighbor nodes of the target node according to different relationship types in the dynamic multi-relationship propagation graph, and output the relationship neighbor set; Step S302: Based on the set of relational neighbors, calculate the association strength between the target node and its neighboring nodes using a relation-specific transformation matrix and a learnable attention vector, and output the relational attention coefficient; Step S303: The initial spatiotemporal feature embeddings of neighboring nodes are weighted and aggregated according to the relational attention coefficients, and iteratively updated by a nonlinear activation function to output the final feature representation of each node.

6. The batch consistency control method based on finished product quality traceability as described in claim 1, characterized in that, Step S40, which involves performing the anomaly root cause localization task based on the final feature representation using a defect node-guided reverse attention tracing method and outputting anomaly root cause probability scores, specifically includes: Step S401: Based on the final feature representation, determine the defective nodes in the finished product quality inspection nodes, and extract the reverse reachable subgraph corresponding to the defective node from the dynamic multi-relationship propagation graph, and output the reverse traceability candidate node set; Step S402: Based on the final feature representation of each candidate node in the reverse traceability candidate node set, calculate the self-abnormal evidence score and reverse attention weight of each candidate node relative to the defective node, and output the root evidence fusion parameters. Step S403: Calculate the anomaly root cause probability score for each candidate node based on the root cause evidence fusion parameters. The calculation formula is as follows: ； in, Candidate nodes The probability score of the root cause of the anomaly For normalization function, The evidence fusion coefficient, Candidate nodes The score of its own abnormal evidence. For a set of relation types, Candidate nodes In relation types The set of downstream neighbor nodes, For downstream nodes Backtracking to candidate nodes Reverse attention weights, For downstream nodes The probability score of the root cause of the anomaly.

7. The batch consistency control method based on finished product quality traceability as described in claim 1, characterized in that, Step S50, which involves generating batch consistency control instructions based on the anomaly root cause probability scores and then outputting the control instructions, specifically includes: Step S501: Generate an anomaly root cause ranking list based on the anomaly root cause probability score of each node, retain the correspondence between node identifier, node type and anomaly root cause probability score, and output the anomaly root cause ranking result. Step S502: Based on a preset probability threshold, filter the ranking results of the abnormal root causes, determine the nodes with abnormal root cause probability scores greater than the preset probability threshold as key abnormal root cause nodes, and output the set of key abnormal root cause nodes. Step S503: Generate corresponding control instructions based on the node type of each key anomaly root source node in the set of key anomaly root source nodes; when the node type is an equipment node, generate equipment early warning and maintenance suggestion instructions; when the node type is a material batch node, generate material isolation and supply chain traceability instructions; when the node type is a process node, generate process parameter adjustment instructions; the control instructions include target node identifier, control action type, execution priority, and control parameters.

8. A batch consistency control system based on finished product quality traceability, applied to the batch consistency control method based on finished product quality traceability as described in any one of claims 1 to 7, characterized in that, The batch consistency control system includes: The dynamic graph construction module is used to acquire batch production traceability data, and based on the batch production traceability data, it performs the task of abstracting the relational relationships of production objects using a dynamic multi-relationship graph construction method, and outputs a dynamic multi-relationship propagation graph. The spatiotemporal coding module is used to perform the node initial representation generation task based on the spatiotemporal coding method constrained by relation type based on the dynamic multi-relation propagation graph, and output the initial spatiotemporal feature embedding. The propagation simulation module is used to perform anomaly propagation representation update task based on the initial spatiotemporal feature embedding using a multi-relation graph attention network, and output the final feature representation; The root cause tracing module is used to perform the abnormal root cause localization task based on the final feature representation using a defect node-guided reverse attention tracing method, and output the abnormal root cause probability score. The control execution module is used to generate batch consistency control instructions based on the probability score of the abnormal root cause and output the control instructions.

9. A batch consistency control device based on finished product quality traceability, characterized in that, The batch consistency control device based on finished product quality traceability includes: a memory, a processor, and a batch consistency control program based on finished product quality traceability stored in the memory and executable on the processor. When the batch consistency control program based on finished product quality traceability is executed by the processor, it implements a batch consistency control method based on finished product quality traceability as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a batch consistency control program based on finished product quality traceability, which, when executed by a processor, implements a batch consistency control method based on finished product quality traceability as described in any one of claims 1 to 7.

Images