An artificial intelligence-based product quality traceability method and system

Abstract

The application discloses a product quality tracing method and system based on artificial intelligence, which comprises the following steps: collecting multi-source quality data, completing correlation mapping processing, constructing quality state vectors of each production window; constructing a reference quality order track, mapping a target product state, and calculating an entropy flow deviation to determine an abnormal window; constructing a time-layered entropy source hypergraph, performing persistent homology calculation, and generating a topological barcode to determine the weight; based on the improved DeepWalk model, adjusting the random walk probability, generating the entropy source fingerprint to determine the dominant entropy source; calculating the input and output and the disturbance quality entropy, identifying the entropy conservation break state, and determining the abnormal propagation state; based on the dominant entropy source and the embedding distance, generating a quality tracing boundary, and outputting a risk set and an evidence chain. The application realizes the identification of latent quality risks, the determination of dominant abnormal sources and the generation of accurate tracing boundaries through quality entropy flow modeling, entropy source topological analysis and improved DeepWalk embedding tracing.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a product quality traceability method and system based on artificial intelligence. Background Technology

[0002] With the development of intelligent manufacturing and industrial informatization, the sources of quality data in the product production process are gradually becoming richer. Enterprises usually record raw material batches, process parameters, equipment status, environmental information, test results and after-sales feedback information of products through MES systems, ERP systems, sensor acquisition systems, testing systems and after-sales management systems. They also establish quality traceability links based on product numbers, batch numbers, process numbers and other identifiers so as to investigate the cause and locate responsibility after product quality problems occur.

[0003] Existing product quality traceability methods primarily rely on explicit correlations for post-event investigations, such as expanding the scope based on the same batch, equipment, process, or raw materials, and combining rule-based judgment, statistical analysis, or conventional machine learning methods to identify anomalies. While these methods can pinpoint some related entities after quality problems have already manifested, they typically use exceeding test results as a trigger condition, making it difficult to detect latent risks where product quality has already deteriorated even before test data exceeds quality assessment thresholds.

[0004] Existing methods typically analyze factors such as equipment, raw materials, processes, and environment as independent traceability objects, making it difficult to characterize the coupling relationship between multiple quality entropy sources in different production windows, and also difficult to distinguish which entropy source mainly causes quality anomalies. Due to the lack of precise determination of the propagation status of quality anomalies and traceability boundaries, existing technologies are prone to the problem of coarse traceability by batch or equipment, resulting in an excessively large range of false recalls, or the omission of potentially risky products from different batches that are affected by the same abnormal entropy source.

[0005] Therefore, how to provide a product quality traceability method and system based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a product quality traceability method and system based on artificial intelligence. This invention comprehensively employs quality entropy flow modeling, time-layered quality entropy source hypergraph analysis, persistent homology topology calculation, and an improved DeepWalk graph embedding method with entropy perturbation constraints to perform correlation analysis on multi-source quality data in the product manufacturing process. This enables the identification of latent anomaly production windows, determination of dominant entropy sources, identification of entropy conservation anomaly violations, and generation of quality traceability boundaries. It has the advantages of strong latent quality risk detection capability, high accuracy in anomaly source identification, accurate traceability range determination, and low false recall rate.

[0007] According to an embodiment of the present invention, a product quality traceability method based on artificial intelligence includes:

[0008] Collect multi-source quality data throughout the entire lifecycle of the target product, perform correlation mapping on the multi-source quality data, and construct a quality status vector of the target product under different production windows;

[0009] A benchmark quality orderliness track is constructed based on the quality state vector of historical qualified products. The quality state vector of the target product is mapped, the deviation of the target product from the benchmark quality orderliness track is calculated, a deviation-type quality entropy flow evolution sequence is generated, and the latent abnormal production window is determined.

[0010] A time-layered quality entropy source hypergraph is constructed based on the latent anomaly production window. Persistent cohomology calculation is performed on the time-layered quality entropy source hypergraph to generate an entropy source coupled topological barcode. The hyperedge entropy perturbation weight is determined based on the high-dimensional feature with the longest persistent lifetime in the topological barcode.

[0011] An improved DeepWalk model with entropy perturbation constraints is used to learn node representations of the time-layered quality entropy source hypergraph. The random walk transition probability is adjusted according to the entropy perturbation weight of the hyperedge to generate an entropy perturbation walk sequence. The quality entropy source fingerprint is generated based on the embedding distance between abnormal product nodes and various entropy source nodes to determine the dominant entropy source corresponding to the target product.

[0012] Based on the dominant entropy source, the input quality entropy, output quality entropy, and disturbance quality entropy of the target product corresponding to the production window are calculated, the entropy conservation anomaly breaking state in the local production window is identified, and the propagation state of quality anomalies in the production process is determined.

[0013] Based on the dominant entropy source, the entropy conservation anomaly breaking state, the embedding distance between the abnormal product node and the product node to be traced, and the quality anomaly propagation state, a product quality traceability boundary is generated, and a set of high-risk products, a set of products with latent risks, a set of products that can be excluded, and the corresponding traceability evidence chain are output.

[0014] Optionally, the multi-source quality data includes raw material data, process parameter data, equipment operation data, environmental parameter data, testing data, product identification data, and after-sales feedback data.

[0015] Optionally, the step of associating and mapping multi-source quality data to construct a quality state vector of the target product under different production windows includes:

[0016] The multi-source quality data is aligned according to product number, work order number, process number and timestamp to obtain the raw material status data, process status data, equipment status data, environmental status data and testing status data of the target product in each production window;

[0017] The raw material status data, process status data, equipment status data, environmental status data, and detection status data are standardized and encoded respectively to obtain raw material status sub-vectors, process status sub-vectors, equipment status sub-vectors, environmental status sub-vectors, and detection status sub-vectors.

[0018] The raw material state sub-vector, process state sub-vector, equipment state sub-vector, environmental state sub-vector, and detection state sub-vector are concatenated according to the production window order to generate the quality state vector of the target product under different production windows.

[0019] Optionally, generating the deviation-type mass entropy flow evolution sequence and determining the latent anomaly production window includes:

[0020] Select historical qualified products from historical production data that have passed final inspection and have no after-sales quality abnormalities. Group the historical qualified products according to the product model, process route, production window division method and inspection node that are consistent with the target product. Extract the quality status vector of each historical qualified product under each production window.

[0021] For each production window, the quality status vectors of multiple historical qualified products belonging to the same group are normalized. The normalized quality status vectors are aggregated to obtain the benchmark quality status vector corresponding to the production window. The benchmark quality status vectors corresponding to each production window are arranged according to the order of process execution to form a benchmark quality order track.

[0022] According to the production window sequence of the target product, the quality state vector of the target product under each production window is compared with the benchmark quality state vector of the same production window in the benchmark quality order track, and the degree of deviation of the target product from the benchmark quality order track under each production window is obtained.

[0023] According to the order of production windows, the degree of state deviation of adjacent production windows is compared. When the degree of state deviation of the later production window is greater than that of the earlier production window, the increase in the deviation between the two is determined as the quality entropy flow deviation of the later production window. The quality entropy flow deviation of each production window is arranged according to the process execution order to generate a deviation-type quality entropy flow evolution sequence.

[0024] Based on the deviation-type quality entropy flow evolution sequence, production windows with continuously increasing quality entropy flow deviation or a single production window with quality entropy flow deviation exceeding a preset deviation threshold are identified. When the detection data corresponding to the production window does not exceed the quality judgment threshold, the production window is determined as a latent abnormal production window.

[0025] Optionally, determining the hyper-edge entropy perturbation weights based on the highest-dimensional features with the longest persistent lifetime in the topological barcode includes:

[0026] Using the latent abnormal production window as the time layer, a time layer sequence is established according to the order of the production windows. In each time layer, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes and abnormal event nodes associated with the deviation event of the target product quality entropy flow are set to obtain a quality entropy source sub-graph arranged in time order.

[0027] In each quality entropy source subgraph, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes that are jointly associated with the same quality entropy flow deviation event are divided into the same entropy source coupling group, and the entropy source coupling group is constructed as a hyperedge in the corresponding time layer.

[0028] Node matching is performed on the quality entropy source subgraphs in two adjacent time layers. The same product node, the same raw material node, the same equipment node, the same process node, the same environment node, the same detection index node, or the same abnormal event node that appear in two adjacent time layers are identified as shared nodes. Cross-layer coupling edges are established based on the quality entropy flow deviation changes of the shared nodes in two adjacent time layers.

[0029] Based on the hyperedges within each time layer and the cross-layer coupling edges between adjacent time layers, a filter complex covering the entropy source coupling relationship within the time layer and the propagation relationship between time layers is constructed. Nodes, hyperedges, and cross-layer coupling edges are added in order of increasing mass entropy flow deviation. Persistent cohomology calculation is performed on the filter complex to record the occurrence position, disappearance position, and topological dimension of the entropy source coupling topological features, and an entropy source coupling topological barcode is generated.

[0030] Based on the entropy source coupled topological barcode, the topological features that have the longest retention time during the continuous filtering process and correspond to the annular entropy source coupled structure or the cavity entropy source coupled structure are selected. The hyperedge associated with the topological features is determined as the main coupled hyperedge. The entropy perturbation weight of the hyperedge is determined according to the degree of deviation of the mass entropy flow, the cross-layer coupling strength and the retention time of the topological features corresponding to the main coupled hyperedge.

[0031] Optionally, the step of generating a quality entropy source fingerprint based on the embedding distance between abnormal product nodes and various entropy source nodes, and determining the dominant entropy source corresponding to the target product, includes:

[0032] Using a time-layered quality entropy source hypergraph as model input data, the input data includes product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, abnormal event nodes, as well as the time layer identifier and hyperedge entropy perturbation weight of each node. An improved DeepWalk model with entropy perturbation constraints is constructed. The improved DeepWalk model includes an input layer, an entropy perturbation weighted adjacency generation layer, a constrained random walk layer, a sequence training layer, a node embedding output layer, and an entropy source determination layer connected in sequence.

[0033] In the entropy perturbation weighted adjacency generation layer, nodes within the same hyperedge are established as adjacency nodes, the entropy perturbation weight of the hyperedge is written into the connection weight between the corresponding adjacency nodes, and the time layer identifier, node type identifier and business object identifier of each node are retained to obtain the entropy perturbation weighted adjacency structure.

[0034] In the constrained random walk layer, each node is used as the starting point of the walk. An entropy perturbation walk sequence is generated according to the connection weight between adjacent nodes. The neighboring node with the larger connection weight is selected as the next walk node with higher priority. The walk length of each starting node, the number of walks corresponding to each node, the number of walk sequences, and the node repetition limit are used as key parameters input into the constrained random walk layer.

[0035] In the sequence training layer, the entropy perturbation walk sequence is used as the sequence training sample, each node in the sequence is used as the center node, and the nodes within a preset window range before and after the center node are used as context nodes. The node representation learning network is trained according to the preset window size, embedding dimension, training rounds and learning rate. The node embedding vector of each node is output in the node embedding output layer.

[0036] In the entropy source determination layer, based on the node embedding vector of the abnormal product node, the embedding proximity of the node embedding vector with the embedding of various entropy source node sets is calculated to generate quality entropy source fingerprints. The entropy source type with the highest embedding proximity and the highest frequency of occurrence in the entropy perturbation walk sequence is determined as the dominant entropy source.

[0037] Optionally, identifying the entropy conservation anomaly violation state in the local production window and determining the propagation state of the quality anomaly in the production process includes:

[0038] Using the latent abnormal production window as the production window to be judged, the quality data corresponding to the dominant entropy source within the production window to be judged is read according to the dominant entropy source.

[0039] Read the quality state vector of the target product when it enters and leaves the production window to be judged, and generate the input quality entropy and the output quality entropy according to the degree of dispersion of each state data in the quality state vector;

[0040] Based on the degree of data fluctuation of the dominant entropy source within the production window to be determined, the degree of contribution of the dominant entropy source to the quality entropy source fingerprint, and the frequency of occurrence of the node corresponding to the dominant entropy source in the entropy perturbation walk sequence, the perturbation quality entropy is generated.

[0041] The difference between the output quality entropy and the input quality entropy is compared, and the increase in output quality entropy is determined by combining the perturbation quality entropy to determine whether it can be explained by the perturbation quality entropy corresponding to the dominant entropy source. If the increase in output quality entropy exceeds the explainable range corresponding to the perturbation quality entropy, the production window to be determined is marked as an entropy conservation anomaly window.

[0042] According to the order of production windows, the continuity of entropy conservation anomaly breaking windows in adjacent production windows is judged. When the entropy conservation anomaly breaking exists only in a single production window and no increase in quality entropy flow deviation occurs in subsequent production windows, the quality anomaly is determined to be in a non-propagating state. When the entropy conservation anomaly breaking occurs continuously in adjacent production windows, the quality anomaly is determined to be in a continuously propagating state. When the entropy conservation anomaly breaking occurs simultaneously in related production windows of different products, different batches, or different process paths, the quality anomaly is determined to be in a diffusion and propagation state.

[0043] Optionally, the output of the high-risk product set, the potentially high-risk product set, the excluded product set, and the corresponding traceability evidence chain includes:

[0044] Taking the dominant entropy source as the starting point for tracing, and selecting the same source node to expand the range according to the node type corresponding to the dominant entropy source;

[0045] Within the extended scope, the product nodes to be traced are classified according to the node embedding distance between the abnormal product nodes and the product nodes to be traced.

[0046] Combining the entropy conservation anomaly breaking state and the quality anomaly propagation state, the high-risk set and the latent risk set are dynamically corrected. When the entropy conservation anomaly breaking state corresponding to the dominant entropy source is diffusion propagation, the adjacent windows before the production window where the dominant entropy source is located are traced back, and product nodes with an embedding distance less than the third distance threshold and located on the diffusion path of the dominant entropy source are added to the high-risk set.

[0047] Based on the product number, batch number, process number, and time window information among the high-risk set, latent risk set, and excludeable set, a product quality traceability boundary is generated, and a traceability evidence chain is output, which includes product identification, risk classification, dominant entropy source type, entropy conservation anomaly breaking state, quality anomaly propagation state, and node embedding distance.

[0048] An artificial intelligence-based product quality traceability system according to an embodiment of the present invention includes:

[0049] The quality status construction module is used to collect and correlate multi-source quality data to generate a quality status vector of the target product under each production window.

[0050] The entropy flow deviation identification module is used to construct a baseline quality order track based on historical qualified products, calculate the quality entropy flow deviation of the target product, and determine the latent abnormal production window.

[0051] The entropy source topology analysis module is used to construct a time-layered quality entropy source hypergraph based on the latent anomaly production window, and to determine the hyperedge entropy perturbation weights through persistent cohomology calculation.

[0052] Entropy perturbation embedding module is used to improve the DeepWalk random walk process based on hyperedge entropy perturbation weights, generate quality entropy source fingerprints and determine the dominant entropy source;

[0053] The entropy conservation determination module is used to calculate the input mass entropy, output mass entropy, and disturbance mass entropy based on the dominant entropy source, and to determine the entropy conservation anomaly violation state and the mass anomaly propagation state.

[0054] The traceability boundary output module is used to output the quality traceability boundary, the set of risky products, and the traceability evidence chain based on the dominant entropy source, the entropy conservation anomaly violation state, the node embedding distance, and the quality anomaly propagation state.

[0055] The beneficial effects of this invention are:

[0056] This invention constructs a baseline quality orderliness track based on historical qualified products and maps the quality state vector of the target product onto this track. This allows for the identification of quality entropy flow deviations before product testing data exceeds the quality judgment threshold, thus identifying potential abnormal production windows in advance. Compared to traceability methods that rely solely on final test results or batch anomaly records, this invention can identify the process of declining product quality orderliness earlier, improving the ability to detect latent quality risks.

[0057] This invention constructs a time-layered quality entropy source hypergraph and combines it with persistent cohomology calculation to obtain entropy source coupling topology barcodes, which can characterize the coupling relationships of products, raw materials, equipment, processes, environment, detection indicators, and abnormal events in different production windows. Simultaneously, this invention uses an improved DeepWalk model with entropy perturbation constraints to generate quality entropy source fingerprints, which can determine the dominant entropy source corresponding to the target product from multiple types of entropy sources, improving the accuracy and interpretability of quality anomaly source identification.

[0058] This invention generates product quality traceability boundaries based on the dominant entropy source, entropy conservation anomaly violation state, node embedding distance, and quality anomaly propagation state, avoiding coarse traceability based solely on the same batch, equipment, or process. This invention can both classify products unaffected by anomaly propagation into an exclusionary product set, reducing the scope of false recalls, and identify potentially risky products from different batches affected by the same anomalous entropy source, improving the completeness and accuracy of quality traceability. Attached Figure Description

[0059] 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:

[0060] Figure 1 This is a flowchart of a product quality traceability method based on artificial intelligence proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the structure of a product quality traceability system based on artificial intelligence proposed in this invention. Detailed Implementation

[0062] 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.

[0063] refer to Figure 1 A product quality traceability method based on artificial intelligence includes:

[0064] Collect multi-source quality data throughout the entire lifecycle of the target product, perform correlation mapping on the multi-source quality data, and construct a quality status vector of the target product under different production windows;

[0065] A benchmark quality orderliness track is constructed based on the quality state vector of historical qualified products. The quality state vector of the target product is mapped, the deviation of the target product from the benchmark quality orderliness track is calculated, a deviation-type quality entropy flow evolution sequence is generated, and the latent abnormal production window is determined.

[0066] A time-layered quality entropy source hypergraph is constructed based on the latent anomaly production window. Persistent cohomology calculation is performed on the time-layered quality entropy source hypergraph to generate an entropy source coupled topological barcode. The hyperedge entropy perturbation weight is determined based on the high-dimensional feature with the longest persistent lifetime in the topological barcode.

[0067] An improved DeepWalk model with entropy perturbation constraints is used to learn node representations of the time-layered quality entropy source hypergraph. The random walk transition probability is adjusted according to the entropy perturbation weight of the hyperedge to generate an entropy perturbation walk sequence. The quality entropy source fingerprint is generated based on the embedding distance between abnormal product nodes and various entropy source nodes to determine the dominant entropy source corresponding to the target product.

[0068] Based on the dominant entropy source, the input quality entropy, output quality entropy, and disturbance quality entropy of the target product corresponding to the production window are calculated, the entropy conservation anomaly breaking state in the local production window is identified, and the propagation state of quality anomalies in the production process is determined.

[0069] Based on the dominant entropy source, the entropy conservation anomaly breaking state, the embedding distance between the abnormal product node and the product node to be traced, and the quality anomaly propagation state, a product quality traceability boundary is generated, and a set of high-risk products, a set of products with latent risks, a set of products that can be excluded, and the corresponding traceability evidence chain are output.

[0070] In this embodiment, the multi-source quality data includes raw material data, process parameter data, equipment operation data, environmental parameter data, testing data, product identification data, and after-sales feedback data.

[0071] In this embodiment, the step of performing correlation mapping on multi-source quality data to construct a quality state vector of the target product under different production windows includes:

[0072] The multi-source quality data is aligned according to product number, work order number, process number and timestamp to obtain the raw material status data, process status data, equipment status data, environmental status data and testing status data of the target product in each production window;

[0073] The raw material status data, process status data, equipment status data, environmental status data, and detection status data are standardized and encoded respectively to obtain raw material status sub-vectors, process status sub-vectors, equipment status sub-vectors, environmental status sub-vectors, and detection status sub-vectors.

[0074] The raw material state sub-vector, process state sub-vector, equipment state sub-vector, environmental state sub-vector, and detection state sub-vector are concatenated according to the production window order to generate the quality state vector of the target product under different production windows.

[0075] In this embodiment, generating the deviation-type mass entropy flow evolution sequence and determining the latent anomaly production window includes:

[0076] Historically qualified products with satisfactory final inspection results and no after-sales quality anomalies are selected from historical production data. These historically qualified products are grouped according to product model, process route, production window division method, and inspection node consistent with the target product. The quality status vector of each historically qualified product under each production window is extracted. Specifically, the extraction of the quality status vector of each historically qualified product under each production window is as follows:

[0077] The system reads the raw material batch number, equipment multi-channel sensor sequence, process parameter sequence, environmental monitoring sequence, and online detection measurement value for each historical qualified product within a specific production window. These five data types are aligned to a unified 100ms sampling grid using millisecond timestamps. The raw material batch number is uniquely encoded to obtain a 12-dimensional raw material state representation. Sixteen statistics are calculated for the equipment sensor sequence: interval mean, variance, maximum value, and kurtosis. Eight statistics are calculated for the process parameter sequence, divided by process segment: interval mean, slope, and fluctuation amplitude. Six statistics are extracted for the environmental monitoring sequence: mean, range, and coefficient of variation. The original values ​​of the online detection measurement values ​​are retained, and four statistics (deviation percentage and Z-score) are added. The resulting 46-dimensional feature vector is normalized to the 0–1 interval and used as the quality state sub-vector for the production window. Each quality state sub-vector is concatenated according to the window number order to form a quality state vector sequence of length 46 × the number of windows.

[0078] For each production window, the quality status vectors of multiple historical qualified products belonging to the same group are normalized. The normalized quality status vectors are then aggregated to obtain the baseline quality status vector corresponding to the production window. These baseline quality status vectors are then arranged according to the order of process execution to form a baseline quality order track. Specifically, the formation of the baseline quality order track is as follows:

[0079] Read the quality status vector of all historical qualified products within each production window. Each quality status vector contains 40 normalized feature components: the first 10 components correspond to raw material indicators, the next 15 to process parameters, the following 10 to equipment operating status, and the last 5 to environmental and online testing results. Average all vectors within the same window on their corresponding components to generate a 40-dimensional window baseline quality status vector. Concatenate the six 40-dimensional baseline vectors end-to-end according to window numbers W01 to W06 to obtain a baseline status sequence of length 240. Then, for adjacent components in the baseline status sequence... The absolute value of the difference between the 40-dimensional vectors of the window is calculated for each component. Components with an absolute value greater than 0.08 are recorded as abrupt components. The abrupt components are smoothed by weighted linear interpolation of the values ​​of the two windows, with the weights distributed according to the production time ratio of adjacent windows. After smoothing, the Euclidean distance between the vectors of adjacent windows is checked again. If the distance is still higher than 0.35, an interpolation window vector is inserted between the two windows, and its components are the average of the smoothed vectors of the two windows. Finally, the window reference vectors and interpolation vectors arranged in sequence are combined into a continuous quality state curve with a length of not less than 240 and not more than 320, which is the reference quality order track.

[0080] Following the production window sequence of the target product, the quality state vector of the target product under each production window is compared item by item with the benchmark quality state vector of the same production window in the benchmark quality ordering orbit. This yields the degree of deviation of the target product's state from the benchmark quality ordering orbit under each production window. Specifically, the degree of deviation of the target product's state from the benchmark quality ordering orbit under each production window is obtained as follows:

[0081] Read the 40-dimensional quality state vector of the target product within a certain production window. Subtract this 40-dimensional vector from the baseline quality state vector within the same window component by component to obtain 40 difference components. Take the absolute value of each of the 40 difference components to generate a 40-dimensional absolute difference vector. Divide the absolute difference vector into five feature categories: raw material indicators, process parameters, equipment status, environmental status, and test results. The raw material indicators correspond to the first 10 components, the process parameters to the next 15 components, the equipment status to the last 10 components, and the environmental status to... For components 36 to 38, the detection results correspond to the last two components; calculate the mean absolute value of the difference within each feature partition to obtain the mean deviation of raw materials, process, equipment, environment, and detection; perform weighted summation of the five deviation means with weights of 0.25, 0.30, 0.25, 0.10, and 0.10 respectively to obtain the window state deviation degree in a single scalar form; repeat the above operation within all production windows to form a state deviation degree sequence with the same number of production windows as the target product;

[0082] According to the order of production windows, the degree of state deviation of adjacent production windows is compared. When the degree of state deviation of the later production window is greater than that of the earlier production window, the increase in the deviation between the two is determined as the quality entropy flow deviation of the later production window. The quality entropy flow deviation of each production window is arranged according to the process execution order to generate a deviation-type quality entropy flow evolution sequence.

[0083] Based on the deviation-type quality entropy flow evolution sequence, production windows with continuously increasing quality entropy flow deviation or where the quality entropy flow deviation of a single production window exceeds a preset deviation threshold are identified. When the detection data corresponding to the production window does not exceed the quality judgment threshold, the production window is designated as a latent abnormal production window. Specifically, the designation of a production window as a latent abnormal production window involves:

[0084] The deviation-type quality entropy flow evolution sequence of the target product is read according to the process sequence. A sliding window check of length 3 is performed on the sequence. If the quality entropy flow deviation of three adjacent production windows is monotonically increasing and the deviation increase of the last window is not less than 1.2 times the increase of the previous window, the last window is marked as a continuously increasing candidate window. The sequence is screened for single-point thresholds. When the quality entropy flow deviation of any window is greater than the preset deviation threshold and at least one of its two adjacent windows has a deviation higher than the 70th percentile of the whole sequence, the window is marked as a threshold-exceeding candidate window. The two types of candidate windows are merged and deduplicated to obtain a candidate latent window set. For each window in the candidate latent window set, the online detection results, finished product detection results, and aging test results of the corresponding window are extracted from the detection database. If the detection results are all higher than their respective quality judgment thresholds, the window is finally confirmed as a latent abnormal production window. If any of the detection results are lower than the judgment threshold, the window is re-recorded as an explicit abnormal window and is no longer included in the latent abnormal range.

[0085] In this embodiment, determining the hyper-edge entropy perturbation weight based on the highest-dimensional feature with the longest persistent lifetime in the topological barcode includes:

[0086] Using latent abnormal production windows as time layers, a time layer sequence is established according to the chronological order of the production windows. Within each time layer, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes associated with the deviation event of the target product's quality entropy flow are set, resulting in a quality entropy source subgraph arranged in chronological order. Specifically, the quality entropy source subgraph arranged in chronological order is as follows:

[0087] Read the window number, start and end time, process number, and target product number of each latent abnormal production window. Arrange the production windows according to the order of process execution and generate a time-layer index. Within each time layer, read the product number, raw material batch number, equipment number, process number, environmental window number, detection index number, and abnormal event number corresponding to the target product, and generate product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes respectively. Write the nodes in the same time layer that have a direct production association with the target product into the node set of the time layer, and establish association edges between product nodes and raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes. Store the node set and association edge set of each time layer in sequence according to the time-layer index to generate a quality entropy source subgraph arranged in time order.

[0088] In each quality entropy source subgraph, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes that are jointly associated with the same quality entropy flow deviation event are divided into the same entropy source coupling group, and the entropy source coupling group is constructed as a hyperedge in the corresponding time layer.

[0089] Node matching is performed on the quality entropy source subgraphs in two adjacent time layers. The same product node, the same raw material node, the same equipment node, the same process node, the same environment node, the same detection index node, or the same abnormal event node that appear in two adjacent time layers are identified as shared nodes. Cross-layer coupling edges are established based on the quality entropy flow deviation changes of the shared nodes in two adjacent time layers.

[0090] Based on the hyperedges within each time layer and the cross-layer coupling edges between adjacent time layers, a filtered complex covering the entropy source coupling relationship within a time layer and the propagation relationship between time layers is constructed. Nodes, hyperedges, and cross-layer coupling edges are added sequentially in order of increasing mass entropy flow deviation. Persistent cohomology calculation is performed on the filtered complex, recording the appearance, disappearance, and topological dimensions of the entropy source coupling topological features, and generating an entropy source coupling topological barcode. Specifically, the construction of the filtered complex covering the entropy source coupling relationship within a time layer and the propagation relationship between time layers involves:

[0091] Read product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, abnormal event nodes, hyperedges, and cross-layer coupling edges within each time layer. Write each node as a zero-dimensional simplex into the initial complex. Convert the association relationship between any two nodes within the same hyperedge into a one-dimensional simplex. Convert the common entropy source coupling relationship formed by any three nodes within the same hyperedge into a two-dimensional simplex. Convert the common entropy source coupling relationship formed by more than three nodes within the same hyperedge into a high-dimensional simplex. Write the cross-layer coupling edge between adjacent time layers as a one-dimensional simplex connecting the complexes of different time layers into the complex structure. Set filter values ​​according to the degree of deviation of mass entropy flow corresponding to nodes, intra-layer simplexes, and cross-layer simplexes, and arrange the filter values ​​from low to high to generate a filtered complex covering the entropy source coupling relationship within the time layer and the propagation relationship between time layers.

[0092] Persistent cohomology calculations are performed on the filtered complex, specifically as follows:

[0093] Nodes, intra-layer simplexes, and cross-layer simplexes are added sequentially in ascending order of filter values. For each node added, the location of the new connected component is recorded. For each one-dimensional simplex added, it is determined whether it connects two existing connected components or forms a closed loop. If it connects two existing connected components, the location where one connected component disappears is recorded; if it forms a closed loop, the location where a one-dimensional topological feature appears is recorded. For each two-dimensional or higher-dimensional simplex added, it is determined whether it fills an existing closed loop or forms a higher-dimensional cavity. If it fills an existing closed loop, the location where the corresponding one-dimensional topological feature disappears is recorded; if it forms a higher-dimensional cavity, the location where the corresponding higher-dimensional topological feature appears is recorded. After traversing all filter values, the location, location, and topological dimension of each topological feature are statistically analyzed. Each topological feature is written in the form of a bar interval to generate an entropy source coupled topological barcode.

[0094] Based on the entropy source coupled topological barcode, the topological features that maintain the longest duration during continuous filtering and correspond to either a ring-shaped or cavity-shaped entropy source coupled structure are selected. The hyperedges associated with these topological features are determined as primary coupled hyperedges. The entropy perturbation weights of the hyperedges are determined based on the deviation of the mass entropy flow, the cross-layer coupling strength, and the retention time of the topological features. Specifically, the determination of the hyperedge entropy perturbation weights is as follows:

[0095] Read the mass entropy flow deviation of each node associated with the main coupling hyperedge, normalize the mass entropy flow deviation of each node and take the average value to obtain the deviation strength value of the main coupling hyperedge; read the number of cross-layer coupling edges corresponding to the main coupling hyperedge in adjacent time layers and the degree of deviation change of each cross-layer coupling edge, normalize and fuse the number of cross-layer coupling edges and the degree of deviation change to obtain the cross-layer coupling strength value of the main coupling hyperedge; read the appearance position and disappearance position of the topological feature in the entropy source coupled topological barcode, and take the interval length between the two as the topological feature retention time; weight and fuse the deviation strength value, cross-layer coupling strength value and topological feature retention time according to preset weights to obtain the hyperedge entropy perturbation weight of the main coupling hyperedge, and write the hyperedge entropy perturbation weight into the corresponding main coupling hyperedge in the time-layered mass entropy source hypergraph.

[0096] In this embodiment, the step of generating a quality entropy source fingerprint based on the embedding distance between abnormal product nodes and various entropy source nodes, and determining the dominant entropy source corresponding to the target product, includes:

[0097] Using a time-layered quality entropy source hypergraph as model input data, the input data includes product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, abnormal event nodes, as well as the time layer identifier and hyperedge entropy perturbation weight of each node. An improved DeepWalk model with entropy perturbation constraints is constructed. The improved DeepWalk model includes an input layer, an entropy perturbation weighted adjacency generation layer, a constrained random walk layer, a sequence training layer, a node embedding output layer, and an entropy source determination layer connected in sequence.

[0098] In the entropy perturbation weighted adjacency generation layer, nodes within the same hyperedge are established as adjacency nodes, the entropy perturbation weight of the hyperedge is written into the connection weight between the corresponding adjacency nodes, and the time layer identifier, node type identifier and business object identifier of each node are retained to obtain the entropy perturbation weighted adjacency structure.

[0099] In the constrained random walk layer, each node is used as the starting point of the walk. An entropy perturbation walk sequence is generated according to the connection weight between adjacent nodes. The neighboring node with the larger connection weight is selected as the next walk node with higher priority. The walk length of each starting node, the number of walks corresponding to each node, the number of walk sequences, and the node repetition limit are used as key parameters input into the constrained random walk layer.

[0100] In the sequence training layer, the entropy perturbation walk sequence is used as the sequence training sample, each node in the sequence is used as the center node, and the nodes within a preset window range before and after the center node are used as context nodes. The node representation learning network is trained according to the preset window size, embedding dimension, training rounds and learning rate. The node embedding vector of each node is output in the node embedding output layer.

[0101] In the entropy source determination layer, based on the node embedding vector of the abnormal product node, the embedding proximity of the node embedding vector with the embedding of various entropy source node sets is calculated to generate a quality entropy source fingerprint. The entropy source type with the highest embedding proximity and the highest frequency of occurrence in the entropy perturbation walk sequence is determined as the dominant entropy source. Specifically, the calculation of the embedding proximity of the node embedding vector with the embedding of various entropy source node sets to generate the quality entropy source fingerprint is as follows:

[0102] Read the embedding vector of the abnormal product node, and extract all embedding vectors from the six types of entropy source node sets: raw materials, equipment, process, environment, detection indicators, and abnormal events. For each type of entropy source node set, compare the position difference between its embedding vector and the embedding vector of the abnormal product node in the vector space, record the difference between the two vectors in each dimension, and accumulate them to obtain the distance between a single node. Select several nodes with the smallest distance in the same entropy source node set as candidate associated nodes, and average the distance values ​​of the candidate associated nodes to obtain the geometric proximity of the entropy source. Count the number of times the candidate associated nodes appear in the entropy perturbation walk sequence, calculate the proportion of the number of times they appear to the total number of times they appear in the same type of node, and obtain the walk activity of this type of entropy source. Perform inverse normalization on the geometric proximity and percentage normalization on the walk activity, and combine the two linearly according to preset weights to obtain the comprehensive proximity of this type of entropy source. Arrange the six comprehensive proximity values ​​in the order of raw materials, equipment, process, environment, detection indicators, and abnormal events to form a six-dimensional quality entropy source fingerprint vector.

[0103] Among them, the entropy perturbation weighted adjacency generation layer includes:

[0104] Hypergraph resolver: Receives the time-layered quality entropy source hypergraph and reads product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, abnormal event nodes, and the time layer identifier of each node;

[0105] Hyperedge expander: expands multiple nodes within the same hyperedge into traversable adjacency relationships;

[0106] Entropy perturbation weight writer: Reads the entropy perturbation weight of each hyperedge and writes the weight into the adjacency relationship obtained by expanding the hyperedge;

[0107] Business Identifier Binder: Binds a node type identifier, production window identifier, product link identifier, and entropy source type identifier to each node;

[0108] Weighted adjacency buffer: Stores an entropy-perturbation-weighted adjacency structure composed of node adjacency relationships, hyperedge entropy perturbation weights, and business identifiers;

[0109] After receiving the time-layered quality entropy source hypergraph, the entropy perturbation weighted adjacency generation layer first reads the node set and hyperedge set in each time layer. For example, in the compression window, it reads product node P, equipment node E, process node G, environment node A, detection index node D, and abnormal event node X. It establishes adjacency relationships between nodes in the same hyperedge that participate in the same quality entropy flow deviation event. For example, it establishes walkable connections between product node P and equipment node E, process node G, and detection index node D. If the entropy perturbation weight of the hyperedge is 0.87, it writes 0.87 into the above adjacency connections. After all hyperedges are expanded, an entropy perturbation weighted adjacency structure is formed, so that random walks no longer perform uniform walks based on ordinary graph connection relationships, but rather perform weighted walks based on the quality entropy flow deviation intensity and entropy source coupling intensity.

[0110] The constrained random walk layer includes:

[0111] Starting point selector: Set the wander starting point according to product node, raw material node, equipment node, process node, environment node, detection index node and abnormal event node respectively;

[0112] Adjacency candidate reader: Reads the set of adjacent nodes of the current node in the entropy perturbation weighted adjacency structure;

[0113] Entropy Perturbation Priority Calculator: Determines the selection priority of the next walking node based on the entropy perturbation weight of the hyperedge in the adjacency relationship;

[0114] Node access controller: Controls the traversal process based on traversal length, number of traversals, and node re-access limits;

[0115] Entropy perturbation sequence buffer: Stores entropy perturbation walk sequences in the order in which nodes are accessed;

[0116] The constrained random walk layer takes an entropy-perturbation-weighted adjacency structure as input and uses the entropy perturbation weights of hyperedges as the random walk constraint rules. It reads all adjacent nodes starting from a given node. For example, if the current node is the abnormal product node P-03842, its adjacent nodes include equipment node E-04, process node G-03, inspection index node D-03, and abnormal event node X-11. If the entropy perturbation weight of the hyperedge from P-03842 to E-04 is higher than the entropy perturbation weight of the hyperedge from P-03842 to raw material node R-07, then equipment node E-04 has a higher priority as the next walk node. The system performs multiple rounds of walks according to a preset walk length, for example, 20 walks per node, 40 nodes per walk sequence, and limits the number of consecutive visits to the same node within the same walk sequence to no more than twice. After the walk is completed, the system generates an entropy perturbation walk sequence containing the propagation relationship of products, equipment, processes, test indicators and abnormal events. This sequence is used to represent the propagation path of the quality entropy flow deviation between different entropy source nodes.

[0117] The sequence training layer includes:

[0118] Sequence Sample Reader: Reads the entropy-perturbed walk sequence output by the constrained random walk layer;

[0119] Central node locator: sequentially sets each node in the walk sequence as the central node;

[0120] Context node sampler: Reads nodes within a preset window range before and after the central node as context nodes;

[0121] Node co-occurrence trainer: Train the node representation learning network based on the co-occurrence relationship between the central node and context nodes;

[0122] Training parameter controller: Sets the context window size, embedding dimension, training epochs, learning rate, and negative sampling count;

[0123] The sequence training layer uses entropy-perturbed walk sequences as training samples and performs representation learning according to the node co-occurrence idea of ​​DeepWalk. It reads an entropy-perturbed walk sequence, such as product node P-03842, equipment node E-04, abnormal event node X-11, process node G-03, detection index node D-03, and product node P-03857, and sequentially designates each node as a center node. When equipment node E-04 is the center node, the system uses product node P-03842, abnormal event node X-11, and process node G-03 within its preceding and following window ranges as context nodes. Since this sequence is generated by entropy perturbed weight constraints, frequently co-occurring nodes indicate a strong correlation in the propagation of quality entropy flow deviation. During training, the system sets the node embedding dimension to 64 or 128, the context window size to 5, the training epochs to 8 to 20, and the learning rate to 0.01 to 0.05, so that product nodes, equipment nodes, process nodes, and abnormal event nodes form comparable structural representations in the vector space.

[0124] The node embedding output layer includes:

[0125] Vector Index Generator: Creates a unique vector index for each product node, raw material node, equipment node, process node, environment node, detection index node, and abnormal event node;

[0126] Embedding vector outputter: Outputs the low-dimensional embedding vector corresponding to each node;

[0127] Node type mapper: binds node embedding vectors to node type identifiers;

[0128] Temporal layer mapper: binds node embedding vectors to production window identifiers;

[0129] Embedding result storage: Stores node embedding results with node type, time layer, and business object identifier;

[0130] The node embedding output layer receives the node representation results trained by the sequence training layer and outputs a corresponding node embedding vector for each node. For example, product node P-03842 outputs a 64-dimensional node embedding vector, equipment node E-04 outputs a 64-dimensional node embedding vector, and abnormal event node X-11 outputs a 64-dimensional node embedding vector. While outputting the vector, the system retains the node type and production window information, enabling subsequent differentiation between product nodes, equipment nodes, process nodes, and inspection indicator nodes. For the same equipment appearing in different production windows, its time-layer identifier is retained, allowing the embedding result of equipment node E-04 in the pressing window W03 to be distinguished from the embedding results of equipment node E-04 in other windows, preventing the mixing of equipment influences from different time states.

[0131] The entropy source determination layer includes:

[0132] Abnormal Product Node Reader: Reads the embedding vector of the abnormal product node to be traced;

[0133] Entropy source node set partitioner: Divides the node embedding vector into raw material node set, equipment node set, process node set, environment node set, detection index node set, and abnormal event node set according to node type;

[0134] Embedding Proximity Calculator: Calculates the embedding proximity between abnormal product nodes and various entropy source node sets;

[0135] Walk frequency counter: Counts the frequency of occurrence of various entropy source nodes in the entropy perturbation walk sequence;

[0136] Entropy Source Fingerprint Generator: Outputs quality entropy source fingerprints according to different entropy source types;

[0137] Dominant entropy source selector: Determines the dominant entropy source based on embedding proximity and walk frequency;

[0138] The entropy source determination layer uses the embedding vector of the abnormal product node as a benchmark, and reads the embedding vectors of various entropy source node sets. For example, the system reads the embedding vector of the abnormal product node P-03842 and calculates its proximity with the raw material node set, equipment node set, process node set, environmental node set, detection index node set, and abnormal event node set, respectively. If P-03842 has the highest embedding proximity with the equipment node set, and the equipment node also has the highest frequency of occurrence in the entropy perturbation walk sequence, then the system determines the equipment entropy source as the dominant entropy source. The quality entropy source fingerprint can be represented as the contribution results of different entropy source types. For example, the raw material entropy source contributes less, the equipment entropy source contributes the most, the process entropy source is second, the environmental entropy source is lower, and the detection index entropy source is used as a feedback correlation term. The final output dominant entropy source is used for entropy conservation anomaly identification and product quality traceability boundary generation.

[0139] In this embodiment, identifying the entropy conservation anomaly violation state in the local production window and determining the propagation state of the quality anomaly in the production process includes:

[0140] Using the latent abnormal production window as the production window to be judged, the quality data corresponding to the dominant entropy source within the production window to be judged is read according to the dominant entropy source.

[0141] Read the quality state vectors of the target product when it enters and leaves the production window to be judged. Generate input and output quality entropy based on the dispersion of each state data in the quality state vector. Specifically, the generation of input and output quality entropy is as follows:

[0142] Based on five feature blocks—raw material characteristics, process parameters, equipment status, environmental parameters, and test results—the quality state vector of the target product is sequentially read when it enters the production window to be judged. Within each feature block, the mean absolute deviation of all components relative to the block mean is calculated to measure the dispersion of the block. The mean absolute deviations of the five feature blocks are combined according to preset weights, with raw material characteristics accounting for 25%, process parameters for 30%, equipment status for 25%, environmental parameters for 10%, and test results for 10%, to obtain the comprehensive dispersion when entering the window. This comprehensive dispersion is recorded as the input quality entropy. The same method is used to read the quality state vector of the target product when it leaves the production window to be judged, and the mean absolute deviations of the five feature blocks are recalculated and combined with the same weights to obtain the comprehensive dispersion when leaving the window. This comprehensive dispersion is recorded as the output quality entropy.

[0143] Based on the data fluctuation level of the dominant entropy source within the production window to be determined, the contribution of the dominant entropy source to the quality entropy source fingerprint, and the frequency of occurrence of the node corresponding to the dominant entropy source in the entropy perturbation walk sequence, a perturbation quality entropy is generated. Specifically, the generation of the perturbation quality entropy is as follows:

[0144] Within the production window to be judged, all quality data components belonging to the dominant entropy source are read, such as pressure, temperature, vibration, current, and equipment alarm information corresponding to the equipment entropy source. The standard deviation of each component is calculated in chronological order, and then these standard deviations are subjected to max-min normalization to obtain the data fluctuation degree. The contribution value of the dominant entropy source is extracted from the quality entropy source fingerprint and normalized by the ratio with the sum of the contribution values ​​of the other five types of entropy sources to obtain the contribution degree. Then, the absolute number of occurrences of the dominant entropy source node is counted from the entropy disturbance walk sequence and normalized by the ratio with the total number of nodes in the walk sequence to obtain the occurrence frequency. Finally, the data fluctuation degree, contribution degree, and occurrence frequency are added according to preset weights, where the weights of data fluctuation degree, contribution degree, and occurrence frequency are set to 0.5, 0.3, and 0.2 respectively. The result is the disturbance quality entropy corresponding to the dominant entropy source.

[0145] The difference between the output quality entropy and the input quality entropy is compared, and the increase in output quality entropy is determined by combining the perturbation quality entropy to determine whether it can be explained by the perturbation quality entropy corresponding to the dominant entropy source. If the increase in output quality entropy exceeds the explainable range corresponding to the perturbation quality entropy, the production window to be determined is marked as an entropy conservation anomaly window.

[0146] According to the order of production windows, the continuity of entropy conservation anomaly breaking windows in adjacent production windows is judged. When the entropy conservation anomaly breaking exists only in a single production window and no increase in quality entropy flow deviation occurs in subsequent production windows, the quality anomaly is determined to be in a non-propagating state. When the entropy conservation anomaly breaking occurs continuously in adjacent production windows, the quality anomaly is determined to be in a continuously propagating state. When the entropy conservation anomaly breaking occurs simultaneously in related production windows of different products, different batches, or different process paths, the quality anomaly is determined to be in a diffusion and propagation state.

[0147] In this embodiment, the output of the high-risk product set, the latent risk product set, the excludeable product set, and the corresponding traceability evidence chain includes:

[0148] Taking the dominant entropy source as the starting point for tracing, and based on the node type corresponding to the dominant entropy source, the same source node expansion range is selected. Specifically, when the dominant entropy source is the equipment entropy source, the product node with the same equipment number and the same time window is selected; when the dominant entropy source is the raw material entropy source, the node with the same raw material batch is selected; when the dominant entropy source is the process entropy source, the node with the same process is selected; and when the dominant entropy source is the environment entropy source, the node with the same environment window is selected.

[0149] Within the extended range, based on the node embedding distance between the abnormal product node and the product node to be traced, the product node to be traced is classified. When the node embedding distance is less than the first distance threshold, the corresponding product node is classified into the high-risk set. When the node embedding distance is between the first distance threshold and the second distance threshold, the corresponding product node is classified into the latent risk set. When the node embedding distance is greater than the second distance threshold, the corresponding product node is classified into the exclude set.

[0150] Combining the entropy conservation anomaly breaking state and the quality anomaly propagation state, the high-risk set and the latent risk set are dynamically corrected. When the entropy conservation anomaly breaking state corresponding to the dominant entropy source is diffusion propagation, the adjacent windows before the production window where the dominant entropy source is located are traced back, and product nodes with an embedding distance less than the third distance threshold and located on the diffusion path of the dominant entropy source are added to the high-risk set.

[0151] Based on the product number, batch number, process number, and time window information among the high-risk set, latent risk set, and excludeable set, a product quality traceability boundary is generated, and a traceability evidence chain is output, which includes product identification, risk classification, dominant entropy source type, entropy conservation anomaly breaking state, quality anomaly propagation state, and node embedding distance.

[0152] refer to Figure 2 An artificial intelligence-based product quality traceability system includes:

[0153] The quality status construction module is used to collect and correlate multi-source quality data to generate a quality status vector of the target product under each production window.

[0154] The entropy flow deviation identification module is used to construct a baseline quality order track based on historical qualified products, calculate the quality entropy flow deviation of the target product, and determine the latent abnormal production window.

[0155] The entropy source topology analysis module is used to construct a time-layered quality entropy source hypergraph based on the latent anomaly production window, and to determine the hyperedge entropy perturbation weights through persistent cohomology calculation.

[0156] Entropy perturbation embedding module is used to improve the DeepWalk random walk process based on hyperedge entropy perturbation weights, generate quality entropy source fingerprints and determine the dominant entropy source;

[0157] The entropy conservation determination module is used to calculate the input mass entropy, output mass entropy, and disturbance mass entropy based on the dominant entropy source, and to determine the entropy conservation anomaly violation state and the mass anomaly propagation state.

[0158] The traceability boundary output module is used to output the quality traceability boundary, the set of risky products, and the traceability evidence chain based on the dominant entropy source, the entropy conservation anomaly violation state, the node embedding distance, and the quality anomaly propagation state.

[0159] Example 1: In a continuous production cycle of a precision electronic module, the system receives multi-source quality data from the production line, involving 8,600 products of the same model. The production process includes six production windows: raw material warehousing, mounting, lamination, curing, online testing, and aging testing. The system collects 18 sets of raw material batch data, approximately 1.26 million pieces of equipment operation data, approximately 920,000 pieces of process parameter data, approximately 340,000 pieces of environmental parameter data, 8,600 sets of online testing data, 8,600 sets of aging test data, and 326 after-sales feedback samples. Each product corresponds to a unique product number, and each production window records the process number, equipment number, environmental window number, and testing node number. There are some inconsistencies in the sampling frequency in the raw data; for example, equipment vibration data is collected every 2 seconds, environmental humidity data is collected every 30 seconds, and online testing data is only generated once after the process is completed.

[0160] After the data enters the processing flow, the system performs correlation mapping on multi-source quality data according to product number, process number, and production window. For example, product P-03842 is associated with raw material batch R-07, pressing equipment E-04, process G-03, environmental window A-02, and inspection index D-03 in pressing window W03. The system encodes 38 fields for this product in W03, including average pressure, pressure fluctuation amplitude, average equipment vibration, ambient humidity, pressing time, and inspection deviation value, into a quality status vector. After data alignment, 51,600 production window quality status vectors were generated for 8,600 products, with the percentage of missing fields decreasing from the original 3.8% to 0.6%, and the outlier sampling point removal rate reaching 1.9%.

[0161] The system selected 3200 historical qualified products of the same model and process route from the historical database, all of which passed final inspection and had no after-sales anomalies. Of these, 2600 were used as the baseline sample and 600 as the validation sample. The system normalized the historical qualified quality state vectors for each production window (W01 to W06) and calculated the stable quality state center for each window, forming a baseline quality order trajectory. After the validation samples were mapped to this trajectory, the average deviations from W01 to W06 were 0.041, 0.052, 0.057, 0.061, 0.064, and 0.069, respectively, indicating that the quality state of qualified products remained stable and changed slowly along the process flow.

[0162] After mapping the target product P-03842 to the baseline quality order track, the system obtained deviations of 0.044, 0.055, 0.186, 0.214, 0.231, and 0.256 for W01 to W06, respectively. The key performance value of this product in the W05 online inspection was 96.7%, higher than the 95% quality judgment threshold; therefore, traditional inspection rules did not trigger an anomaly. However, the method of this invention found that the deviation of the quality entropy flow of P-03842 from W02 to W03 increased from 0.011 to 0.131, exceeding the preset deviation threshold of 0.080, and continued to increase from W04 to W06. Therefore, the system identified the W03 pressing window as a latent anomaly production window.

[0163] The system constructs a time-layered quality entropy source hypergraph based on this latent anomaly production window. The W03 time layer contains 8600 product nodes, 18 raw material nodes, 12 equipment nodes, 6 process nodes, 9 environmental nodes, 24 detection index nodes, and 31 abnormal event nodes. Taking P-03842 as an example, its quality entropy flow deviation event is simultaneously associated with product node P-03842, equipment node E-04, process node G-03, environmental node A-02, detection index node D-03, and abnormal event node X-11. The system forms a hyperedge around these nodes. The entire time-layered quality entropy source hypergraph forms a total of 1246 hyperedges and 3928 cross-time-layer shared node coupling relationships.

[0164] The system performs persistent cohomology calculations on the time-layered quality entropy source hypergraph. During the process of adding nodes, hyperedges, and cross-layer coupling edges at each level, the system records the appearance and disappearance of entropy source coupling topological features, generating entropy source coupling topological barcodes. The calculation results show that the high-dimensional feature with the longest persistent lifetime corresponds to hyperedge H-287. This hyperedge mainly connects equipment E-04, pressing process G-03, vibration anomaly event X-11, and 43 product nodes, with a persistent lifetime of 0.74, higher than the average persistent lifetime of 0.29 for other high-dimensional features. Based on this, the system identifies H-287 as the main coupling hyperedge and sets its hyperedge entropy perturbation weight to 0.87.

[0165] In the improved DeepWalk model, the system transforms the temporally hierarchical quality entropy source hypergraph into an entropy-perturbation-weighted walkable structure. Nodes within the same hyperedge are designated as adjacent nodes, and the entropy perturbation weights of the hyperedge serve as random walk constraints. The model is configured with a walk length of 40, 20 walks per node, a node embedding dimension of 64, a context window size of 5, and 8 training epochs. The system generates a total of 24,860 entropy-perturbation walk sequences, of which sequences passing through the combined paths E-04, X-11, and G-03 account for 31.6%, while this path only appears in 9.4% of the ordinary DeepWalk model.

[0166] After node embedding training, the system uses the abnormal product node P-03842 as a benchmark to calculate its embedding proximity to various entropy source node sets. The calculation results show that P-03842 has a proximity of 0.86 with the equipment entropy source node set, appearing 382 times in the entropy perturbation walk sequence; a proximity of 0.63 with the process entropy source node set, appearing 211 times; a proximity of 0.28 with the raw material entropy source node set, appearing 74 times; a proximity of 0.24 with the environmental entropy source node set, appearing 61 times; and a proximity of 0.47 with the detection index node set, appearing 136 times. The quality entropy source fingerprint generated by the system shows that equipment entropy sources are dominant; therefore, the pressing equipment E-04 is identified as the dominant entropy source corresponding to the target product.

[0167] During the entropy conservation anomaly violation identification phase, the system reads the mass state vectors of P-03842 before and after entering W03. The input mass entropy of this product before entering W03 is 0.42, and the output mass entropy after leaving W03 is 0.71. The disturbance mass entropy corresponding to device E-04 is 0.16. The system determines that 0.13 of the increase in output mass entropy relative to input mass entropy cannot be explained by the device disturbance mass entropy, exceeding the violation threshold of 0.08. Therefore, W03 is identified as an entropy conservation anomaly violation window. Further examination of adjacent products reveals that among the 146 products processed by device E-04 within the same operating window, 119 showed an increase in mass entropy flow deviation after W03. Of these, 36 belong to different batches. Based on this, the system classifies this anomaly as a diffusion propagation state.

[0168] The system ultimately generates a traceability boundary based on the dominant entropy source, node embedding distance, entropy conservation anomaly violation state, and propagation state. The results show a high-risk product set of 97 items, a latent risk product set of 42 items, and a product set that can be excluded of 1117 items. High-risk products are mainly concentrated within the same pressing condition window of equipment E-04, while latent risk products include 36 products from different batches but from the same equipment and following the same abnormal event path. The traceability evidence chain output by the system includes product number, production window, equipment node, main coupling hyperedge, hyperedge entropy perturbation weight, quality entropy source fingerprint, entropy conservation anomaly violation state, and propagation state.

[0169] In the comparative experiment, the traditional traceability method included 1256 products in the re-inspection according to the rules of the same batch and the same equipment. Ultimately, 103 products with actual risks were identified, 1153 products were mistakenly included in the re-inspection, and 29 products with risks from different batches were missed. The traceability boundary generation took approximately 42 minutes. The method of this invention included 139 products in the re-inspection, identified 128 products with actual risks, mistakenly included 11 products in the re-inspection, and missed 4 products with risks. The traceability boundary generation took approximately 8 minutes. The accuracy rate of locating latent anomaly windows using the traditional method was 62.4%, while that of the method of this invention was 91.7%. The accuracy rate of identifying the dominant anomaly source using the traditional method was 57.9%, while that of the method of this invention was 88.6%. The re-inspection workload of the traditional method was 1256 products, while that of the method of this invention was 139 products, reducing the re-inspection workload by approximately 88.9%.

[0170] 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 product quality traceability method based on artificial intelligence, characterized in that, include: Collect multi-source quality data throughout the entire lifecycle of the target product, perform correlation mapping on the multi-source quality data, and construct a quality status vector of the target product under different production windows; A benchmark quality orderliness track is constructed based on the quality state vector of historical qualified products. The quality state vector of the target product is mapped, the deviation of the target product from the benchmark quality orderliness track is calculated, a deviation-type quality entropy flow evolution sequence is generated, and the latent abnormal production window is determined. A time-layered quality entropy source hypergraph is constructed based on the latent anomaly production window. Persistent cohomology calculation is performed on the time-layered quality entropy source hypergraph to generate an entropy source coupled topological barcode. The hyperedge entropy perturbation weight is determined based on the high-dimensional feature with the longest persistent lifetime in the topological barcode. An improved DeepWalk model with entropy perturbation constraints is used to learn node representations of the time-layered quality entropy source hypergraph. The random walk transition probability is adjusted according to the entropy perturbation weight of the hyperedge to generate an entropy perturbation walk sequence. The quality entropy source fingerprint is generated based on the embedding distance between abnormal product nodes and various entropy source nodes to determine the dominant entropy source corresponding to the target product. Based on the dominant entropy source, the input quality entropy, output quality entropy, and disturbance quality entropy of the target product corresponding to the production window are calculated, the entropy conservation anomaly breaking state in the local production window is identified, and the propagation state of quality anomalies in the production process is determined. Based on the dominant entropy source, the entropy conservation anomaly breaking state, the embedding distance between the abnormal product node and the product node to be traced, and the quality anomaly propagation state, a product quality traceability boundary is generated, and a set of high-risk products, a set of products with latent risks, a set of products that can be excluded, and the corresponding traceability evidence chain are output.

2. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The multi-source quality data includes raw material data, process parameter data, equipment operation data, environmental parameter data, testing data, product identification data, and after-sales feedback data.

3. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The step of associating and mapping multi-source quality data to construct a quality state vector of the target product under different production windows includes: The multi-source quality data is aligned according to product number, work order number, process number and timestamp to obtain the raw material status data, process status data, equipment status data, environmental status data and testing status data of the target product in each production window; The raw material status data, process status data, equipment status data, environmental status data, and detection status data are standardized and encoded respectively to obtain raw material status sub-vectors, process status sub-vectors, equipment status sub-vectors, environmental status sub-vectors, and detection status sub-vectors. The raw material state sub-vector, process state sub-vector, equipment state sub-vector, environmental state sub-vector, and detection state sub-vector are concatenated according to the production window order to generate the quality state vector of the target product under different production windows.

4. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The generation of the deviation-type quality entropy flow evolution sequence and the determination of the latent anomaly production window include: Select historical qualified products from historical production data that have passed final inspection and have no after-sales quality abnormalities. Group the historical qualified products according to the product model, process route, production window division method and inspection node that are consistent with the target product. Extract the quality status vector of each historical qualified product under each production window. For each production window, the quality status vectors of multiple historical qualified products belonging to the same group are normalized. The normalized quality status vectors are aggregated to obtain the benchmark quality status vector corresponding to the production window. The benchmark quality status vectors corresponding to each production window are arranged according to the order of process execution to form a benchmark quality order track. According to the production window sequence of the target product, the quality state vector of the target product under each production window is compared with the benchmark quality state vector of the same production window in the benchmark quality order track, and the degree of deviation of the target product from the benchmark quality order track under each production window is obtained. According to the order of production windows, the degree of state deviation of adjacent production windows is compared. When the degree of state deviation of the later production window is greater than that of the earlier production window, the increase in the deviation between the two is determined as the quality entropy flow deviation of the later production window. The quality entropy flow deviation of each production window is arranged according to the process execution order to generate a deviation-type quality entropy flow evolution sequence. Based on the deviation-type quality entropy flow evolution sequence, production windows with continuously increasing quality entropy flow deviation or a single production window with quality entropy flow deviation exceeding a preset deviation threshold are identified. When the detection data corresponding to the production window does not exceed the quality judgment threshold, the production window is determined as a latent abnormal production window.

5. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The step of determining the hyper-edge entropy perturbation weights based on the highest-dimensional features with the longest persistent lifetime in the topological barcode includes: Using the latent abnormal production window as the time layer, a time layer sequence is established according to the order of the production windows. In each time layer, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes and abnormal event nodes associated with the deviation event of the target product quality entropy flow are set to obtain a quality entropy source sub-graph arranged in time order. In each quality entropy source subgraph, product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, and abnormal event nodes that are jointly associated with the same quality entropy flow deviation event are divided into the same entropy source coupling group, and the entropy source coupling group is constructed as a hyperedge in the corresponding time layer. Node matching is performed on the quality entropy source subgraphs in two adjacent time layers. The same product node, the same raw material node, the same equipment node, the same process node, the same environment node, the same detection index node, or the same abnormal event node that appear in two adjacent time layers are identified as shared nodes. Cross-layer coupling edges are established based on the quality entropy flow deviation changes of the shared nodes in two adjacent time layers. Based on the hyperedges within each time layer and the cross-layer coupling edges between adjacent time layers, a filter complex covering the entropy source coupling relationship within the time layer and the propagation relationship between time layers is constructed. Nodes, hyperedges, and cross-layer coupling edges are added in order of increasing mass entropy flow deviation. Persistent cohomology calculation is performed on the filter complex to record the occurrence position, disappearance position, and topological dimension of the entropy source coupling topological features, and an entropy source coupling topological barcode is generated. Based on the entropy source coupled topological barcode, the topological features that have the longest retention time during the continuous filtering process and correspond to the annular entropy source coupled structure or the cavity entropy source coupled structure are selected. The hyperedge associated with the topological features is determined as the main coupled hyperedge. The entropy perturbation weight of the hyperedge is determined according to the degree of deviation of the mass entropy flow, the cross-layer coupling strength and the retention time of the topological features corresponding to the main coupled hyperedge.

6. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The process of generating a quality entropy source fingerprint based on the embedding distance between abnormal product nodes and various entropy source nodes, and determining the dominant entropy source corresponding to the target product, includes: Using a time-layered quality entropy source hypergraph as model input data, the input data includes product nodes, raw material nodes, equipment nodes, process nodes, environmental nodes, detection index nodes, abnormal event nodes, as well as the time layer identifier and hyperedge entropy perturbation weight of each node. An improved DeepWalk model with entropy perturbation constraints is constructed. The improved DeepWalk model includes an input layer, an entropy perturbation weighted adjacency generation layer, a constrained random walk layer, a sequence training layer, a node embedding output layer, and an entropy source determination layer connected in sequence. In the entropy perturbation weighted adjacency generation layer, nodes within the same hyperedge are established as adjacency nodes, the entropy perturbation weight of the hyperedge is written into the connection weight between the corresponding adjacency nodes, and the time layer identifier, node type identifier and business object identifier of each node are retained to obtain the entropy perturbation weighted adjacency structure. In the constrained random walk layer, each node is used as the starting point of the walk. An entropy perturbation walk sequence is generated according to the connection weight between adjacent nodes. The neighboring node with the larger connection weight is selected as the next walk node with higher priority. The walk length of each starting node, the number of walks corresponding to each node, the number of walk sequences, and the node repetition limit are used as key parameters input into the constrained random walk layer. In the sequence training layer, the entropy perturbation walk sequence is used as the sequence training sample, each node in the sequence is used as the center node, and the nodes within a preset window range before and after the center node are used as context nodes. The node representation learning network is trained according to the preset window size, embedding dimension, training rounds and learning rate. The node embedding vector of each node is output in the node embedding output layer. In the entropy source determination layer, based on the node embedding vector of the abnormal product node, the embedding proximity of the node embedding vector with the embedding of various entropy source node sets is calculated to generate quality entropy source fingerprints. The entropy source type with the highest embedding proximity and the highest frequency of occurrence in the entropy perturbation walk sequence is determined as the dominant entropy source.

7. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The process of identifying entropy conservation anomaly violations in local production windows and determining the propagation status of quality anomalies during production includes: Using the latent abnormal production window as the production window to be judged, the quality data corresponding to the dominant entropy source within the production window to be judged is read according to the dominant entropy source. Read the quality state vector of the target product when it enters and leaves the production window to be judged, and generate the input quality entropy and the output quality entropy according to the degree of dispersion of each state data in the quality state vector; Based on the degree of data fluctuation of the dominant entropy source within the production window to be determined, the degree of contribution of the dominant entropy source to the quality entropy source fingerprint, and the frequency of occurrence of the node corresponding to the dominant entropy source in the entropy perturbation walk sequence, the perturbation quality entropy is generated. The difference between the output quality entropy and the input quality entropy is compared, and the increase in output quality entropy is determined by combining the perturbation quality entropy to determine whether it can be explained by the perturbation quality entropy corresponding to the dominant entropy source. If the increase in output quality entropy exceeds the explainable range corresponding to the perturbation quality entropy, the production window to be determined is marked as an entropy conservation anomaly window. According to the order of production windows, the continuity of entropy conservation anomaly breaking windows in adjacent production windows is judged. When the entropy conservation anomaly breaking exists only in a single production window and no increase in quality entropy flow deviation occurs in subsequent production windows, the quality anomaly is determined to be in a non-propagating state. When the entropy conservation anomaly breaking occurs continuously in adjacent production windows, the quality anomaly is determined to be in a continuously propagating state. When the entropy conservation anomaly breaking occurs simultaneously in related production windows of different products, different batches, or different process paths, the quality anomaly is determined to be in a diffusion and propagation state.

8. The product quality traceability method based on artificial intelligence according to claim 1, characterized in that, The output of the high-risk product set, the latent risk product set, the excludeable product set, and the corresponding traceability evidence chain includes: Taking the dominant entropy source as the starting point for tracing, and selecting the same source node to expand the range according to the node type corresponding to the dominant entropy source; Within the extended range, the product nodes to be traced are classified according to the node embedding distance between the abnormal product nodes and the product nodes to be traced. Combining the entropy conservation anomaly breaking state and the quality anomaly propagation state, the high-risk set and the latent risk set are dynamically corrected. When the entropy conservation anomaly breaking state corresponding to the dominant entropy source is diffusion propagation, the adjacent windows before the production window where the dominant entropy source is located are traced back, and product nodes with an embedding distance less than the third distance threshold and located on the diffusion path of the dominant entropy source are added to the high-risk set. Based on the product number, batch number, process number, and time window information among the high-risk set, latent risk set, and excludeable set, a product quality traceability boundary is generated, and a traceability evidence chain is output, which includes product identification, risk classification, dominant entropy source type, entropy conservation anomaly breaking state, quality anomaly propagation state, and node embedding distance.

9. A product quality traceability system based on artificial intelligence, comprising executing the product quality traceability method based on artificial intelligence as described in any one of claims 1 to 8, characterized in that, include: The quality status construction module is used to collect and correlate multi-source quality data to generate a quality status vector of the target product under each production window. The entropy flow deviation identification module is used to construct a baseline quality order track based on historical qualified products, calculate the quality entropy flow deviation of the target product, and determine the latent abnormal production window. The entropy source topology analysis module is used to construct a time-layered quality entropy source hypergraph based on the latent anomaly production window, and to determine the hyperedge entropy perturbation weights through persistent cohomology calculation. Entropy perturbation embedding module is used to improve the DeepWalk random walk process based on hyperedge entropy perturbation weights, generate quality entropy source fingerprints and determine the dominant entropy source; The entropy conservation determination module is used to calculate the input mass entropy, output mass entropy, and disturbance mass entropy based on the dominant entropy source, and to determine the entropy conservation anomaly violation state and the mass anomaly propagation state. The traceability boundary output module is used to output the quality traceability boundary, the set of risky products, and the traceability evidence chain based on the dominant entropy source, the entropy conservation anomaly violation state, the node embedding distance, and the quality anomaly propagation state.

Images