Shoe processing production process analysis system based on industrial big data

Abstract

The application relates to the field of industrial big data and intelligent manufacturing technology, in particular to a shoe processing production process analysis system based on industrial big data, which comprises a data acquisition module, which acquires equipment time sequence data, environmental sensing data and unstructured process logs of discrete manufacturing links such as cutting, sewing and forming, maps the feature vectors into graph database entities after vectorization and establishes initial correlation topological weights; a context monitoring module, which extracts production context features and calculates deviation from a historical good product context baseline, and outputs a context drift state; a dynamic topological reconstruction module, which performs space-time lag alignment based on cross-process timestamp difference when drift occurs, updates the correlation topological weights between entities and generates a reconstructed graph; and an adaptive retrieval module, which receives an abnormal query instruction, performs cross-modal joint retrieval in combination with the current context and the reconstructed graph, and outputs matched cross-modal causal relationship data; the application can realize high correlation retrieval and root cause tracing of shoe processing abnormalities.

Description

Technical Field

[0001] This invention relates to the field of industrial big data and intelligent manufacturing technology, specifically to a shoe processing production process analysis system based on industrial big data. Background Technology

[0002] Shoe manufacturing is a typical discrete manufacturing scenario, with its production chain typically encompassing multiple process stages such as cutting, sewing, and molding. To ensure product quality stability and traceability of anomalies, it is usually necessary to uniformly manage and analyze multi-source data generated during production, including equipment timing data, workshop environment data, and process logs. As shoe factories become increasingly digitalized, the types of data collected on-site continue to increase, and the data sources become more dispersed. In order to take into account the flow relationships between different processes and the needs of anomaly investigation, it is usually necessary to establish correlations between various types of production data based on preset process flows, and to conduct production status analysis and anomaly cause retrieval accordingly.

[0003] However, when analyzing the discrete manufacturing process of shoe processing, it is necessary to address the problem of data association across processes, multiple types, and asynchronous arrivals. Existing methods often establish data connections according to fixed process paths or static topologies, assuming a direct and stable correspondence between upstream data and downstream results, or using methods such as matching at the same time or single threshold filtering for anomaly analysis. This approach is insufficient to fully reflect the impact of changes in production context, such as material batch variations, temperature and humidity fluctuations, and glue replacements, on quality results. At the same time, there is usually a significant time lag between cutting, sewing, and molding. Relying solely on static rules or single data modalities for retrieval can easily lead to inaccurate identification of cross-process causal chains, resulting in low efficiency in anomaly root cause localization, and thus affecting the reliability and effectiveness of shoe processing production process analysis. Summary of the Invention

[0004] The purpose of this invention is to provide a shoe manufacturing process analysis system based on industrial big data, and to solve the following technical problems:

[0005] Organizing originally scattered, asynchronous, and difficult-to-compare multi-source heterogeneous data into a graph structure that can dynamically evolve with the production context avoids reconstruction bias caused by static time matching, and enables more efficient and accurate cross-modal joint retrieval and root cause tracing of complex shoe processing anomalies.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A shoe manufacturing process analysis system based on industrial big data, the system comprising:

[0008] The data acquisition module is configured to collect multi-source heterogeneous data from multiple process steps in the discrete manufacturing process of shoe processing, and to convert the multi-source heterogeneous data into feature vectors and unify them to the same dimension through linear projection or zero-padding operations, and then map them into data entities in a preset graph database. Based on the preset shoe processing process flow rules, the module establishes the initial association topology weights between the data entities.

[0009] The context monitoring module is configured to extract production context features from the multi-source heterogeneous data in real time, obtain the deviation value between the production context features and the context baseline pre-constructed based on historical good product data by calculating the feature distance, and output the context drift status.

[0010] The dynamic topology reconstruction module is configured to receive the context drift state. When the context drift state indicates that a context drift has occurred, it extracts the timestamp difference between data across process links as a spatiotemporal lag parameter and aligns the multi-source heterogeneous data. Based on the aligned multi-source heterogeneous data, it updates the association topology weights between data entities across process links in the graph database and generates a reconstructed graph.

[0011] The adaptive retrieval module is configured to receive anomaly query instructions through a human-computer interaction interface, perform cross-modal joint retrieval based on the current production context features and the reconstructed graph, and extract and output cross-modal causal relationship data that matches the anomaly query instructions;

[0012] The multi-source heterogeneous data includes equipment time-series data, environmental sensor data, and unstructured process logs.

[0013] Optionally, the context monitoring module includes:

[0014] The feature extraction unit is configured to extract the current batch material attribute features and workshop microclimate features from the multi-source heterogeneous data as the production context features;

[0015] The state determination unit is configured to: determine that context drift has occurred when the deviation value is greater than a preset drift threshold, and output a first context drift state; and output a second context drift state and maintain the current associated topology weight when the deviation value is less than or equal to the preset drift threshold.

[0016] Optional, the dynamic topology reconfiguration module includes:

[0017] The spatiotemporal alignment unit is configured to align the device timing data and the unstructured process log based on the spatiotemporal lag parameter, and generate an aligned data sequence.

[0018] The weight update unit is configured to update the associated topological weights using the aligned data sequence to generate the reconstructed graph;

[0019] The cross-process steps include the cutting step, the sewing step, and the forming step.

[0020] Optional, the adaptive retrieval module includes:

[0021] The parsing unit is configured to parse the anomaly query instruction to obtain the target anomaly feature vector;

[0022] The association calculation unit is configured to obtain the contextual association degree between the target anomaly feature vector and each of the data entities by calculating the similarity between the target anomaly feature vector and the feature vectors of each of the data entities in the reconstructed graph.

[0023] The feature output unit is configured to: extract the corresponding data entity as the cross-modal causal relationship data output when the context relevance is greater than a preset relevance threshold; and add an irrelevant label to the corresponding data entity and filter it when the context relevance is less than or equal to the preset relevance threshold.

[0024] Optionally, the system may also include:

[0025] The isolated data mounting module is configured to compare the data source network identifier of the input data stream with a preset data access whitelist to detect undefined detached data streams, extract the metadata features and spatiotemporal context features of the detached data streams, and vectorize and concatenate the metadata features and spatiotemporal context features. By calculating the similarity between the concatenated feature vector and the feature vectors of each data entity in the graph database, the matching probability between the detached data stream and each data entity in the reconstructed graph is obtained.

[0026] Optionally, the isolated data mounting module can also be configured as follows:

[0027] When the matching probability is greater than the preset matching threshold, the free data stream is used as a new attribute graph node and attached to the data entity with the highest matching probability in the reconstructed graph.

[0028] When the matching probability is less than or equal to the preset matching threshold, the free data stream is stored in the preset isolated storage area and a manual review prompt is generated.

[0029] Optionally, the system may also include:

[0030] The index maintenance module is configured to periodically calculate the weighted sum of the change in query time and the change in retrieval hit rate of the adaptive retrieval module in the current period, as the joint query performance decay rate;

[0031] When the performance decay rate of the joint query is greater than the preset decay threshold, a local index reconstruction instruction is triggered to reconstruct the index of active subgraphs in the reconstructed graph whose update frequency is greater than the preset frequency threshold, while maintaining the original index of subgraphs whose update frequency is less than or equal to the preset frequency threshold.

[0032] When the performance decay rate of the joint query is less than or equal to the preset decay threshold, the current index structure is maintained.

[0033] Optionally, the material attribute characteristics of the current batch include the batch replacement record of the shoe upper material;

[0034] The workshop microclimate characteristics include workshop temperature and humidity fluctuation data;

[0035] The cross-modal causal relationship data includes cross-modal characteristics that characterize abnormal debonding of the molding process due to changes in adhesive curing time caused by fluctuations in workshop temperature and humidity.

[0036] Optionally, the data acquisition module is deployed on an edge computing node;

[0037] The graph database is deployed on a cloud server;

[0038] The adaptive retrieval module receives abnormal query commands through a human-computer interaction interface and displays cross-modal causal relationship data.

[0039] The beneficial effects of this invention are:

[0040] 1) This invention monitors production context features such as material properties and microclimate in real time. When context drift is detected, it introduces the timestamp difference across process links as a spatiotemporal lag parameter for dynamic topology reconstruction. This mechanism breaks the traditional static connection based on fixed paths, effectively overcomes the misalignment problem of upstream data acting asynchronously on downstream in discrete manufacturing, and significantly improves the accuracy of the graph in representing the real production relationship.

[0041] 2) This invention performs cross-modal joint retrieval based on the current context and reconstructed graph. It is not limited to a single data modality, but comprehensively evaluates the similarity and correlation of device data, environmental sensing, and text logs. By automatically filtering weakly related entities, the system can output a coherent cross-modal causal chain, transforming the originally scattered data stack into a highly explanatory correlation path, which greatly improves the efficiency of anomaly tracing and location.

[0042] 3) This invention designs an island data mounting mechanism for undefined data streams such as newly added sensors or temporary fields; it calculates the matching probability by extracting metadata and spatiotemporal context features, and automatically realizes the mounting or isolation review of graph nodes based on thresholds; this enables the system to smoothly absorb new process evidence without polluting the existing graph structure, and has adaptive expansion and evolution capabilities.

[0043] 4) To address the issue of query performance degradation caused by the continuous evolution of the graph during the production process, this invention introduces a performance degradation monitoring mechanism. By calculating the combined degradation rate of query time and hit rate, the system only triggers local index reconstruction for frequently updated active subgraphs when necessary. This avoids resource consumption and online business blockage caused by full reconstruction, and effectively coordinates the consumption of computing resources with the online retrieval efficiency of dynamic graphs.

[0044] 5) This invention adopts an architecture that coordinates edge nodes and cloud servers. High-frequency raw data is preprocessed into feature vectors at the edge and then aggregated to the cloud for centralized management of the graph and complex cross-modal retrieval. This allocation method greatly reduces the risk of high-frequency transmission congestion in the factory network, ensuring the real-time nature of on-site data collection and early warning, while leveraging the computing power advantage of the cloud for cross-batch and in-depth causal analysis. Attached Figure Description

[0045] The invention will now be further described with reference to the accompanying drawings.

[0046] Figure 1 This is a schematic diagram of the modules of the shoe processing production process analysis system based on industrial big data provided in the embodiments of this application. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 The shoe manufacturing process analysis system based on industrial big data includes: a data acquisition module, configured to collect multi-source heterogeneous data from multiple process links in the discrete manufacturing process of shoe manufacturing, and to convert the multi-source heterogeneous data into feature vectors and unify them to the same dimension through linear projection or zero-padding operations, and then map them into data entities in a preset graph database. The system establishes the initial association topology weights between each data entity based on the preset shoe manufacturing process flow rules.

[0049] The context monitoring module is configured to extract production context features from multi-source heterogeneous data in real time, obtain the deviation value between the production context features and the context baseline pre-built based on historical good product data by calculating the feature distance, and output the context drift status.

[0050] The dynamic topology reconstruction module is configured to receive context drift status. When the context drift status indicates that context drift has occurred, it extracts the timestamp difference between data across process links as a spatiotemporal lag parameter and aligns the multi-source heterogeneous data. Based on the aligned multi-source heterogeneous data, it updates the association topology weights between data entities across process links in the graph database and generates a reconstructed graph.

[0051] The adaptive retrieval module is configured to receive anomaly query commands through a human-machine interface, perform cross-modal joint retrieval based on the current production context features and reconstructed graph, and extract and output cross-modal causal relationship data that matches the anomaly query command; among which, multi-source heterogeneous data includes equipment time series data, environmental sensor data and unstructured process logs.

[0052] This embodiment provides a production process analysis mechanism for discrete manufacturing scenarios in shoe processing. Specifically, based on the continuous flow process of the same order in a shoe factory in the three core processes of cutting, sewing, and molding, the mechanism maps equipment time sequence data, environmental sensor data, and unstructured process logs into an evolvable data map, and dynamically corrects the data associations across processes when the production context changes, thereby supporting the joint retrieval of the causes of anomalies.

[0053] To facilitate the explanation of the technical solution, this embodiment introduces a context baseline vector. With the current batch context vector To characterize the production status, and using nodes. , , Characterizing spectral entities, in The symbols represent the initial or updated edge weights between entities; the above symbols are only used in the example deduction of this embodiment and do not change the meaning of each module, data entity and associated topology weight in the embodiment.

[0054] During the data acquisition phase, the system collects at least three types of data from each workstation. The first type is equipment time-series data, such as the cutting machine pressure curve, sewing machine motor speed, and forming press holding time series. The second type is environmental sensor data, such as the cutting workshop temperature, sewing machine workshop humidity, and forming area dew point value.

[0055] The third category is unstructured process logs, such as text records filled in by the team leader at 10:00 AM, such as temporarily adjusting the suture tension at the sewing machine station when changing the B27 batch of glue; to facilitate unified processing, the system first performs feature vectorization on different types of data.

[0056] One feasible approach is to extract the mean, peak value, fluctuation amplitude, and duration from the device time-series data to form a numerical vector of length 4; and to extract the current value, the average value over the past 30 minutes, and the slope of change from the environmental sensor data to form a numerical vector of length 3.

[0057] The process logs are first segmented into words, and then keywords such as glue replacement, rework, and insufficient pressing are encoded into discrete tags, which are further mapped into semantic vectors of length 4. Then, through zero padding or linear projection, the data from different sources are uniformly transformed into the same dimension, for example, uniformly transformed into entity feature vectors of length 6. The linear projection is specifically implemented through a pre-trained fully connected neural network layer or principal component analysis algorithm, so as to map the data to the same dimension vector space while preserving the multidimensional feature variance of data from different sources. The fully connected neural network layer is pre-trained based on historical shoe processing heterogeneous data samples.

[0058] Linear projection is specifically implemented through pre-trained fully connected neural network layers or principal component analysis algorithms to map data from different sources to a vector space of the same dimension while preserving the multidimensional feature variances of the data.

[0059] During the graph database mapping phase, each vectorized record is written into the graph database as a data entity node; nodes can carry attributes according to process, time period, batch number, and equipment number.

[0060] The initial association topology weights are established based on the preset shoe processing flow rules; for example, if the same shoe upper component is first cut and then transferred to the sewing machine to enter the forming process, then a directed edge is established in the graph database for the cutting node → sewing machine node → forming node.

[0061] The initial weight of an edge can be assigned according to the process adjacency relationship. For example, the initial weight of directly adjacent processes is set to 0.8, and the initial weight of edges that are not directly adjacent but have a work order inheritance relationship is set to 0.4.

[0062] To illustrate this process, a simplified calculation example is given; assume that there are only three entities in a batch: the pruning node. Sewing machine node Molding nodes ;in, The vector is , The vector is , The vector is ;

[0063] Based on the process flow rules, the system is established and Two edges, with initial weights of 0.8 and 0.8 respectively; if there is a historical experience path in the same order where abnormal cutting humidity affected molding delamination, then an additional path is established. The edges can be initially set to 0.4; in this way, a computable initial topology is constructed.

[0064] During the context monitoring phase, the system extracts production context features in real time and compares them with the context baseline built based on historical good products; here, the context is not a single sensor value, but a reflection of the combined state of the current production environment and process conditions.

[0065] Historical good product baselines can be obtained through statistical analysis of multiple normal batch samples. For example, the baseline vector can be obtained by encoding the material type, glue batch, temperature and humidity range, and equipment load range in several historical good product batches. The distance between the current production context vector and the baseline vector can be any one of Euclidean distance, cosine distance, or weighted Manhattan distance.

[0066] For example, let the historical good product baseline vector be... The current batch context vector is If Euclidean distance is used, the deviation value is obtained by taking the square root of the sum of squares of the four-dimensional differences, and the calculated result is approximately 0.52.

[0067] To simplify the explanation, the sum of the absolute values ​​of the deviations in each dimension can also be used. This is only intended as a visual illustration of a large degree of deviation. In actual deployment, the drift determination should use the same distance metric and the drift threshold should be set accordingly to avoid inconsistencies between the illustrated algorithm and the actual determination rules.

[0068] If the system presets the drift threshold to 0.5 under the Euclidean distance caliber, then the current deviation value of 0.52 indicates that the production context has significantly deviated from the historical good product state, and therefore outputs a state signal indicating that context drift has occurred.

[0069] During the dynamic topology reconfiguration phase, the system receives the aforementioned status signals and, upon detecting drift, extracts the timestamp difference between data across process stages as a spatiotemporal lag parameter.

[0070] Shoe manufacturing is a typical discrete manufacturing process. Data generated by upstream processes does not immediately affect downstream anomalies, but may appear in the forming stage after several hours. Therefore, the system does not use a simple same-time matching method, but introduces spatiotemporal lag alignment. Here, time in spatiotemporal space corresponds to the timestamp difference between processes, and space corresponds to the process, workstation, or equipment location label.

[0071] The former is used to calculate the lag length, while the latter is used to limit alignment to only within the same order, the same part, or the same process path, thereby avoiding the mis-attaching of records from different workstations and different production objects into the same causal chain.

[0072] For example, if the cutting process records a material batch replacement at 09:10, the sewing process starts processing the batch of parts at 11:00, and the molding process experiences a degumming anomaly at 14:30, then the system extracts the time difference from cutting to sewing (110 minutes) and the time difference from sewing to molding (210 minutes), and uses these as the spatiotemporal lag parameters on the corresponding path of this batch.

[0073] Then, the system uses the part number, work order number, or batch label as anchor points to re-align data entities that are separated by different time intervals; after alignment, the edge weights in the graph database are updated according to the actual strength of the association.

[0074] To demonstrate how edge weights are updated, we will continue using the three nodes mentioned above. , , Assume the original edge weights are , , After alignment, it was found that: Increased humidity and The degumming abnormality repeatedly occurred in multiple batches, and the similarity increased from 0.35 to 0.78 after time-delay alignment. It can be increased from 0.4 to 0.75;

[0075] like If the correlation between the vibration signal from the sewing machine in the middle section and the current anomaly weakens, then... The value can be reduced from 0.8 to 0.55; the reconstructed graph is more reflective of the current actual production relations, rather than remaining on the static connection of a fixed process flow.

[0076] During the adaptive retrieval phase, the system receives anomaly query instructions through the human-computer interaction interface and performs cross-modal joint retrieval based on the current context and the reconstructed graph. Cross-modal means that the same anomaly is not searched from only one type of data, but simultaneously in numerical sensing, time-series operation records and text logs to find evidence of causal relationship.

[0077] For example, a process engineer inputs a query to determine the cause of molding debonding; the system first parses the query into a target anomaly feature vector. , then calculate The similarity to the vectors of each entity in the graph is calculated, and the similarity is further weighted in light of the current context.

[0078] If the current context indicates that this batch is in a high humidity environment, paths related to entities such as humidity fluctuations in adhesive curing and shortened holding time will be given higher search priority;

[0079] The final output data is not just about the abnormal parameters of a molding machine, but also provides cross-modal causal chains. For example, the decrease in the holding time of the molding press at 13:50 is related to the increase in workshop humidity from 58% to 72% at 10:20, which in turn is related to the log record of replacing the B27 glue batch at 09:10. It also points out that these entities are connected into an abnormal root cause path through the updated association weights.

[0080] It should be noted that when a certain type of data is missing at the current moment, the system does not terminate the processing directly; if the equipment timing data is interrupted for a short time, the environmental data and process logs can be retained to participate in the map update, and a low confidence mark can be set for the missing edges.

[0081] If unstructured logs cannot be effectively parsed, the original text and timestamps are retained for later supplementary recording or offline parsing; if the current batch cannot match any historical good product baseline, the baseline of similar material batches in the recent period is used as a temporary reference, and an uncertainty measurement mark is added to this drift determination.

[0082] If all correlations after reconstruction are lower than the output threshold, the system will return a message to the human-computer interaction interface indicating that no highly reliable causal chain was found, along with a list of the top-ranked entities for manual review.

[0083] A shoe factory changed the batch of upper material and glue in the morning while producing a batch of winter sports shoes; the equipment curve of the cutting section did not exceed the preset safety threshold range, but the humidity in the workshop exceeded the preset standard humidity tolerance range compared with the historical good product period; the sewing section experienced fluctuations in sewing rhythm at noon, and the molding section experienced local delamination in the afternoon.

[0084] The system first vectorizes the equipment data, temperature and humidity data, and team logs of the three processes to construct a data map corresponding to the order; it identifies that the current batch context deviates from the good product baseline, thus triggering topology reconstruction; after considering the time lag between cutting, sewing and molding, the system increases the cross-process edge weight of the high humidity + new glue batch, and when the engineer initiates a query for the cause of glue delamination, it outputs a cross-modal causal path pointing to the combined effect of humidity change and glue curing time change;

[0085] The purpose of this step is to organize the originally scattered, asynchronous, and difficult-to-compare multi-source heterogeneous data into a reconfigurable graph structure, and to enable the correlation between data to evolve dynamically with changes in the production context, thereby achieving highly relevant retrieval and root cause tracing of complex shoe processing anomalies.

[0086] In a preferred embodiment of the present invention, the context monitoring module includes: a feature extraction unit configured to extract the current batch material attribute features and workshop microclimate features from multi-source heterogeneous data as production context features; and a state determination unit configured to: determine that context drift has occurred when the deviation value is greater than a preset drift threshold, and output a first context drift state; and output a second context drift state and maintain the current associated topology weight when the deviation value is less than or equal to the preset drift threshold.

[0087] This embodiment provides a more specific context monitoring mechanism. Specifically, based on the above, if the context is only understood as general environmental changes, it is difficult to distinguish which type of change will actually affect the shoe processing quality, which can easily lead to invalid reconstruction. Therefore, this embodiment further limits the context features to the material properties of the current batch and the microclimate of the workshop, so that the drift judgment is more in line with the real process sensitive factors of shoe processing.

[0088] Material properties primarily reflect the differences in the response of the same process parameter to different materials; these properties may include the type of upper material, material batch number, glue batch, lining type, and whether there are temporary substitute materials, etc.; for ease of unified calculation, each item can be coded as a numerical value; for example, leather is coded as 0.9, and mesh is coded as 0.4.

[0089] The original planned batch and the alternative batch are coded as 0 and 1 respectively; if the upper material in an order is switched from the original batch to the alternative batch, the feature value of this dimension will change abruptly, and the system will regard it as key context information that may affect the stability of subsequent processes.

[0090] Workshop microclimate characteristics are used to reflect the potential impact of the local environment on the processing status of adhesives, materials, and equipment; these characteristics may include temperature, humidity, dew point, and the amplitude of temperature and humidity fluctuations over the past 30 minutes;

[0091] Instead of using only a single measurement, this system uses statistics within a short time window to avoid false triggering by instantaneous noise. For example, if the current humidity is 72%, the average humidity over the past 30 minutes is 69%, and the fluctuation range is 8%, the system generates a microclimate sub-vector that includes the current value, the average value, and the degree of fluctuation.

[0092] The system concatenates material property characteristics and workshop microclimate characteristics into a production context vector for the current batch. For ease of explanation, a simplified vector of length 5 can be constructed to represent the material type, whether it replaces a batch, the glue batch code, the average temperature, and the humidity fluctuation value in sequence.

[0093] Assuming the historical baseline for good products is The current batch is If weighted distance calculation is used, and high weights are given to whether the batch is replaced and humidity fluctuation value, such as 0.3 and 0.25 respectively, while the weights of other dimensions are low, the final deviation value will increase significantly. Assuming the calculation result is 0.68, and the system preset drift threshold is 0.50, the state determination unit will output the first context drift state.

[0094] If in another batch, the current vector is If the difference from the baseline is small, assuming the deviation is only 0.12, the system outputs the second context drift state and keeps the association weights in the existing graph unchanged. The reason for this is that if the context does not change substantially, frequent reconstruction will only bring unnecessary computational overhead and may also introduce short-term deviations.

[0095] Furthermore, when material attribute characteristics are missing, such as when the supply chain system fails to synchronize new batch labels in a timely manner, the system can first use the planned material code of the current work order as a temporary material value and add a low confidence identifier to this dimension.

[0096] When the microclimate sensor malfunctions, it can read the sensor value of the adjacent workstation or use the statistical value of the most recent effective time window for short-term replacement. However, if the replacement time exceeds the preset limit, such as 30 minutes, the automatic drift judgment will be suspended and manual confirmation will be prompted.

[0097] If the deviation value is exactly equal to the threshold, this embodiment follows a conservative strategy of not triggering reconstruction, that is, outputting the second context drift state to reduce the risk of false triggering.

[0098] In the aforementioned winter sports shoe order, at 9:00 AM, the supply chain system recorded that the shoe upper was switched from the original leather batch to another inventory batch, and at the same time, the humidity in the molding workshop rose rapidly due to weather changes.

[0099] After the system extracts the two features of batch replacement of shoe upper material and increased humidity fluctuation, it calculates that the deviation value of the current batch compared with the historical good products exceeds the threshold, and then marks it as the first context drift state;

[0100] If another batch of the same style of shoes in the afternoon only has slight temperature fluctuations and the materials have not been changed, then output the second state, continue to use the existing associated topology, and do not enter the reconstruction process;

[0101] The purpose of this step is to explicitly extract the contextual factors that are truly strongly related to fluctuations in shoe processing quality, and to distinguish between changes that need to be reconstructed and disturbances that do not require action through thresholding, thereby achieving a more stable and targeted contextual awareness.

[0102] In a preferred embodiment of the present invention, the dynamic topology reconstruction module includes: a spatiotemporal alignment unit configured to align device timing data and unstructured process logs based on spatiotemporal hysteresis parameters, thereby generating an aligned data sequence;

[0103] The weight update unit is configured to update the associated topological weights using the aligned data sequence to generate a reconstructed graph; the cross-process stages include the cutting stage, the sewing stage, and the forming stage.

[0104] This embodiment provides a topology reconstruction mechanism for cross-process time delay problems. Specifically, in the aforementioned scheme, although reconstruction can be triggered after detecting context drift, if data from different processes are spliced ​​together only according to absolute time proximity, it is easy to mistake irrelevant records for causal chains, which is especially obvious when there is a long flow between cutting, sewing and forming.

[0105] Therefore, this embodiment introduces a spatiotemporal alignment unit and a weight update unit to specifically handle the asynchronous correspondence between equipment timing data and unstructured process logs;

[0106] The spatiotemporal alignment unit extracts key timestamps from the three stages of cutting, sewing, and molding for the same order, the same part, or the same batch. Key timestamps can include cutting completion time, sewing line entry time, molding entry time, or event occurrence time in text logs, such as 09:10 glue replacement, 11:20 manual glue reapplication, etc. The system uses the work order number or part tracking code as the primary key to link data from different sources to the same production object.

[0107] The system calculates the timestamp difference between adjacent processes to form a time-space lag parameter. For example, if a shoe upper part is cut at 09:00, goes on the sewing machine at 11:00, and enters the forming machine at 14:20, then the time-space lag parameter from cutting to sewing is 120 minutes, and the time-space lag parameter from sewing to forming is 200 minutes.

[0108] If the log records a change of glue batch at 09:10, the potential lag between this log and the molding anomaly at 14:20 is 310 minutes.

[0109] During alignment, the system does not simply take records with exactly the same time, but establishes an alignment window centered on the lag parameter; the time width of the alignment window is dynamically set based on the historical average processing tolerance time of the corresponding process, the material buffer time, or the data sampling period of the equipment sensor.

[0110] For example, for the abnormal event at 14:20, backtrack the log event about 310 minutes ago and the sewing machine timing segment about 200 minutes ago;

[0111] If the equipment timing data is sampled by minute and the log is recorded by event time point, then the 14:20 molding anomaly can be mapped forward to the sewing timing segment around 11:00 and the cutting or process log event around 09:10; thus, an aligned data sequence is obtained.

[0112] To facilitate understanding, a very simple example of an aligned sequence is constructed; consider a forming anomaly node. Occurred at 14:20, sewing machine vibration timing segment node Covering 10:55 to 11:05, process log node Record 09:10: Replace B27 glue; After calculating the hysteresis parameters, the system believes... and There is a 200-minute correspondence, with There is a 310-minute correspondence; thus, an alignment sequence is formed. ;

[0113] If multiple subsequent anomalies repeat with the same sequence, it indicates that the sequence has stable explanatory power across the process path; it should be noted that the above-mentioned forming anomaly nodes used for local extrapolation Sewing machine timing segment node and process log nodes This is for illustrative purposes only and does not correspond to or replace the context vector symbols mentioned above.

[0114] The weight update unit updates the associated topological weights in the graph using the aligned data sequence; a simplified implementation is to consider the initial process weight, the number of co-occurrences after alignment, and the vector similarity for each edge simultaneously;

[0115] For example, the new association topological weights are calculated as a weighted sum of the decayed values ​​of the old weights and the new weights calculated based on alignment evidence; assuming the edges The original weight was 0.4. Statistical analysis revealed that 7 out of 10 abnormal batches showed an aligned sequence of glue change logs and molding delamination, with a vector similarity of 0.8. Therefore, the edge... It can be increased to 0.76;

[0116] If the edge If it appears only 3 times out of 10 anomalous batches, with a similarity of 0.5, then the edge... It can be updated to 0.58; in this way, the generated reconstructed map will no longer only reflect the sequence of processes, but also the actual correlation strength after spatiotemporal lag correction;

[0117] Furthermore, if the same molding anomaly corresponds to multiple possible upstream events, such as two adjacent glue replacement log times being close, the system can retain the top two candidate paths and assign them normalized confidence scores, such as 0.55 and 0.45 respectively.

[0118] If an upstream process lacks a usable timestamp, the edges related to that process will not be enhanced in this round of reconstruction; only the original weights will be retained. If the length of the sequence obtained after alignment is insufficient, for example, only the data of the forming end is found but there is no corresponding record in the upstream, the system will mark the anomaly as a local link, which will only be used for retrieval reference and not as a high-confidence causal path.

[0119] In the aforementioned order, the cutting process completed a batch of shoe upper pieces at 9:00 AM, the team log recorded the batch of glue replacement at 9:10 AM, the sewing process processed this batch of parts at around 11:00 AM, and the molding process began to experience concentrated glue detachment at 2:20 PM.

[0120] The spatiotemporal alignment unit connects these three time points by component number to generate an alignment sequence of glue change record - sewing machine processing - molding anomaly. After the weight update, the edge weight of glue batch change → glue delamination anomaly exceeds the edge weight of sewing machine vibration → glue delamination anomaly. Therefore, in subsequent searches, the system prioritizes glue-related paths as candidate root causes.

[0121] The purpose of this step is to explicitly incorporate the time misalignment of processes in discrete manufacturing into the data organization process, avoiding reconstruction deviations caused by static time matching, thereby achieving a more accurate map reconstruction that conforms to the actual causal propagation laws of production.

[0122] In a preferred embodiment of the present invention, the adaptive retrieval module includes: a parsing unit configured to parse an anomaly query instruction to obtain a target anomaly feature vector; and an association calculation unit configured to obtain the contextual association degree between the target anomaly feature vector and each data entity by calculating the similarity between the target anomaly feature vector and the feature vectors of each data entity in the reconstructed graph.

[0123] The feature output unit is configured to: extract the corresponding data entity as cross-modal causal relationship data output when the context relevance is greater than the preset relevance threshold; and add irrelevant tags to the corresponding data entity and filter it when the context relevance is less than or equal to the preset relevance threshold.

[0124] This embodiment provides a more refined adaptive retrieval mechanism; specifically, after the aforementioned reconstructed graph is generated, if there is a lack of parsing and filtering rules for abnormal queries, the system may output a large number of related but useless entities, causing engineers to still need to manually check them one by one.

[0125] Therefore, this embodiment uses a parsing unit, an association calculation unit, and a feature output unit to vectorize the query intent and filter highly reliable entities in the current context;

[0126] The parsing unit is used to convert the anomaly query command received by the human-computer interaction interface into a target anomaly feature vector; the query command can be structured input or short text;

[0127] For example, querying the cause of molding delamination abnormalities can retrieve relevant records of molding delamination after sewing; the parsing unit can first extract keywords and then map them to a preset abnormal semantic space;

[0128] Specifically, the mapping process uses a pre-trained industrial domain word vector model or a Transformer-based language model to convert the extracted keywords into fixed-dimensional word vectors. The word vector model or the Transformer-based language model is pre-trained based on a historical shoe processing anomaly diagnosis corpus.

[0129] For ease of explanation, let's assume that molding debonding corresponds to the basic semantic vector. The high humidity environment context correction vector is Then the target anomaly feature vector can be formed as ;

[0130] The association computation unit traverses each data entity in the reconstructed graph, calculating its eigenvectors and... The similarity between them is then weighted according to the current production context to obtain the contextual relevance.

[0131] The contextual relevance here is not limited to pure vector similarity, but can comprehensively consider the semantic similarity between the entity and the query, the edge weight of the path in which the entity is located, and whether the current batch context supports the path;

[0132] Construct a simplified computational example; assume there are three candidate entities in the graph: entity This indicates that B27 glue was replaced at 09:10, and the vector is... ;entity This indicates that the sewing machine vibration exceeded the limit at 11:00, and the vector is... ;entity This indicates that the humidity rose to 72% at 10:20, and the vector is... ;

[0133] If we only look at vector similarity and The similarity can be set to 0.88. It is 0.43. The value is 0.97; considering the material replacement and high humidity environment in the current batch context, the system can add context weighting to entities related to humidity and adhesive curing, for example... The weighted correlation coefficient is 0.91. The weighted average is 0.99, while Because it is weakly related to the current context, the correlation remains at 0.45; if the preset correlation threshold is 0.70, then... , The output is preserved. Irrelevant tags were added and the filter was applied;

[0134] In addition to outputting a single entity, the feature output unit can also output a causal chain composed of multiple highly correlated entities; taking the above example again, if and There is an edge weight of 0.82 between them. If there is an edge weight of 0.87 between the molded and debonded entities, the system can output cross-modal causal relationship data of increased humidity → batch replacement of glue → molded debonding anomaly, instead of displaying two isolated records separately; this is more conducive to engineers understanding the anomaly propagation chain.

[0135] It should be noted that when the query text is too short or the expression is ambiguous, such as only the reason for the exception is entered, the parsing unit can call the context of the most recent exception work order as a supplementary constraint; if the correlation degree of multiple entities is higher than the threshold but there is no connection between them, the system can divide them into multiple candidate causal clusters and display them separately.

[0136] If the relevance of all entities is below the threshold, the feature output unit returns an empty result and indicates that no highly relevant entities were found. It also includes the top few entities with the highest scores but below the threshold for manual review. For entities marked as irrelevant, an expiration period can be set, such as filtering within the current query session, to avoid permanently excluding potentially useful data in the future.

[0137] After the aforementioned debonding incident occurred, the process engineer entered the query for the cause of the debonding anomaly in this batch through the terminal; the parsing unit converted the query into a target vector of debonding + high humidity + glue curing anomaly;

[0138] After correlation calculation, the system found that the humidity rise record and the B27 glue replacement log had a higher correlation with the target anomaly than the threshold. Although the sewing machine vibration exceeded the limit for a short time, the correlation was insufficient and therefore it was automatically filtered out. The final interface only outputs two highly correlated data and their path relationship for engineers to check first.

[0139] The purpose of this step is to reduce the massive number of entities in the graph to a small number of evidence nodes that are most relevant to the current anomaly, and to automatically filter out weakly relevant interference items, thereby achieving more efficient and accurate cross-modal anomaly retrieval.

[0140] In a preferred embodiment of the present invention, the system further includes: an isolated data mounting module, configured to detect undefined detached data streams by comparing the data source network identifier of the input data stream with a preset data access whitelist, extracting metadata features and spatiotemporal context features of the detached data streams, vectorizing and concatenating the metadata features and spatiotemporal context features, and obtaining the matching probability between the detached data streams and each data entity in the reconstructed graph by calculating the similarity between the concatenated feature vector and the feature vectors of each data entity in the graph database.

[0141] This embodiment provides a mechanism for mounting isolated data for new data sources. Specifically, in the aforementioned scheme, the existing map can effectively organize known data types, but when shoe factories continuously introduce new equipment, new sensors, or new record tables, there are often isolated data streams where data has been collected but access rules have not been predefined.

[0142] Directly discarding data would lead to the loss of potential clues; forcibly accessing all data would easily pollute the existing data map. Therefore, this embodiment introduces an isolated data mounting module to perform feature extraction and matching probability evaluation on undefined data streams.

[0143] The isolated data mounting module detects the free data stream; the free data stream here can be the output of a newly installed glue viscosity sensor, or it can be a note field stream temporarily added by the quality inspection terminal.

[0144] The detection process is as follows: the data source network identifier and protocol header features of the input data stream are compared in real time through the message parsing protocol. When it is found that the network identifier is not included in the preset data access whitelist, or the parsed field name and data type do not match the local preset device object model template, the input data stream is determined to be an undefined free data stream.

[0145] Subsequently, the system extracts the metadata features and spatiotemporal context features of the free data stream. The metadata features mainly reflect the identity information of the data itself, such as the source device number, field name, sampling frequency, unit type, and workstation.

[0146] The spatiotemporal context features mainly reflect the time interval in which it appears, the process in which it occurs, the corresponding work order, and the environmental status of nearby sensors. To facilitate unified comparison, the system vectorizes the two types of features and concatenates them into a comprehensive vector.

[0147] For example, a newly connected glue viscosity sensor outputs data every 5 minutes and is installed at the glue application station before molding; the system can extract metadata sub-vectors from it, including the pre-molding station code (assigned a value of 0.8), sensor class label (assigned a value of 0.9), and sampling frequency (assigned a value of 0.2). Then, extract the spatiotemporal context sub-vectors containing the high humidity state (assigned a value of 0.7), the order batch matching degree (assigned a value of 1.0), and the time distance from the degumming event (assigned a value of 0.8). The combined feature vector is obtained after concatenation. ;

[0148] System Calculation The similarity between the entity vectors and existing entity vectors in the reconstructed graph is converted into matching probabilities. For ease of explanation, it is assumed that there are three entities in the graph: entity... The log for changing glue batches has a vector similarity of 0.86; entity... For recording the molding temperature and humidity, the similarity is 0.82; solid model. The similarity to the sewing machine vibration record is 0.34.

[0149] The system can be further normalized to obtain the matching probability, for example It is 0.46. It is 0.43. The value is 0.11; therefore, this free data stream is more likely to belong to the glue process related to pre-molding rather than the sewing equipment process.

[0150] The reason for needing this mechanism is that the aforementioned retrieval and reconstruction mainly revolve around existing nodes. When new sensors or temporary new fields appear, without a pre-mounting process, the new evidence cannot enter the existing causal chain, and therefore cannot reflect its data representation role in cross-modal causal analysis in subsequent retrievals.

[0151] Furthermore, when the feature loss rate of the detached data stream at the metadata level exceeds the preset security threshold, such as missing source device number and timestamp, and only retaining the numerical ontology, the system can first use the most basic time proximity and work order correlation for coarse matching;

[0152] If even the work order number is missing, it will only be isolated and cached, and will not participate in the current graph matching; if its matching probability with multiple entities is close, for example, the highest is 0.41 and the second highest is 0.39, the system can retain the multiple candidate state and wait for more context to arrive before making the final mounting.

[0153] During the aforementioned investigation into order anomalies, the shoe factory temporarily activated a new type of glue viscosity sensor, but this sensor was not previously included in the standard data access template.

[0154] After the system identifies it as a free data stream, it extracts information such as originating from the pre-molding glue application station, sampling frequency of 5 minutes, overlap with the current de-glue order time, and being in a high humidity environment, and calculates the similarity between it and existing map entities;

[0155] The results showed that it had a high matching probability with two entities: glue batch change and abnormal molding humidity, and was therefore judged to be strongly coupled with the debonding path.

[0156] The purpose of this step is to enable new, undefined data streams to obtain preliminary attribution judgments without waiting for manual remodeling, thereby achieving the map's adaptive expansion capability for new data types.

[0157] In a preferred embodiment of the present invention, the isolated data mounting module is further configured to: when the matching probability is greater than a preset matching threshold, mount the detached data stream as a new attribute graph node to the data entity with the highest matching probability in the reconstructed graph; when the matching probability is less than or equal to the preset matching threshold, store the detached data stream in a preset isolated storage area and generate a manual review prompt.

[0158] This embodiment provides a decision mechanism for mounting free data; specifically, after the aforementioned matching probability calculation, if there are no clear mounting and isolation rules, the system will still face two risks: one is to forcibly incorporate low-confidence data into the graph, causing an incorrect causal chain;

[0159] Secondly, new data that potentially contains effective process features has long been isolated nodes and has failed to be connected to the reconstructed graph network, thus failing to participate in retrieval. Therefore, this embodiment introduces threshold-based decision-making to divert the detached data to two paths: automatic mounting and isolated review.

[0160] The system first reads the matching probability of each candidate entity in the free data stream and selects the maximum value as the basis for mounting judgment. If the maximum value is greater than the preset matching threshold, it is determined that the free data stream has a sufficiently high degree of belonging to the target entity and can be mounted as a new attribute graph node to the corresponding entity.

[0161] The new attribute graph node can be understood as either an independent node or an extended attribute node belonging to the original entity. Both can retain their original sampled values, timestamps, and source identifiers.

[0162] For example, continuing with the aforementioned novel adhesive viscosity sensor data, if it is consistent with the physical... The maximum matching probability of the glue batch change log is 0.46, with the entity. The probability of recording the molding temperature and humidity is 0.43;

[0163] Assuming the system's automatic mounting threshold is set to 0.45, if this detached data stream exceeds the threshold, the system will treat it as a new node. Mount to Up, and establish or The associated edge has the attribute set to "Mount Source = Isolated Data Module Mount Probability = 0.46".

[0164] The aforementioned new nodes This is used solely as a code for newly added map nodes to describe the mounting result, unlike its use as a vector symbol in the aforementioned embodiments. Irrelevant;

[0165] Conversely, if the highest matching probability of a certain free data stream is only 0.28, which does not exceed the threshold, the system will not incorporate it into the formal map, but will write it into a preset isolated storage area; the isolated storage area will still retain the original data, extracted features, candidate matching results and time information for subsequent manual review.

[0166] At the same time, the system generates a manual review prompt, such as finding an unassigned data stream with a maximum matching probability of 0.28, suggesting checking the field source or adding a new process template;

[0167] The reason for adopting this mechanism is that the previous layer only completed the preliminary calculation of feature vector similarity, but the high similarity does not naturally mean that it can be directly added to the image; only by setting a threshold boundary can a balance be achieved between automation and reliability.

[0168] Furthermore, when the highest matching probability and the second highest matching probability are too close, for example, 0.47 and 0.46 respectively, although the highest value is greater than the threshold, the system can add a low-discrimination label and send the mounting result to the manual review queue at the same time.

[0169] If the same detached data stream is isolated and stored multiple times in a short period of time, and subsequently frequently shows a high degree of similarity with the same entity group, the system can accumulate evidence and upgrade it to a candidate standard data source after reaching the cumulative number threshold.

[0170] If the capacity of the isolated storage area reaches the preset limit, priority will be given to retaining recent data with high suspected value, while data that has not been reviewed for a long time and has a very low probability of matching will be archived and compressed.

[0171] In the above-mentioned investigation of adhesive delamination anomalies, the matching probability between the newly added adhesive viscosity data and the adhesive batch change log was 0.46, which is slightly higher than the threshold of 0.45. Therefore, the system automatically attached the data to the adhesive process link.

[0172] Another set of temporary notes from the handheld terminal only contained the words "abnormal feel" and lacked a clear correspondence between workstation and time. Its highest matching probability was only 0.22, so it was stored in the isolation area and a review prompt was sent to the process engineer. After manual confirmation, it can be decided whether the notes data should be officially incorporated into the atlas.

[0173] The purpose of this step is to provide a controlled access mechanism for new and unmodeled data, thereby reducing the detrimental effect of non-physical artifacts on the quality of the maps while preserving the ability to absorb evidence of new processes.

[0174] In a preferred embodiment of the present invention, the system further includes: an index maintenance module, configured to periodically calculate the weighted sum of the change in query time and the change in retrieval hit rate of the adaptive retrieval module in the current period, as the joint query performance decay rate;

[0175] When the performance decay rate of the join query exceeds the preset decay threshold, a local index reconstruction instruction is triggered to reconstruct the indexes of active subgraphs in the reconstructed graph whose update frequency exceeds the preset frequency threshold, while maintaining the original indexes of subgraphs whose update frequency is less than or equal to the preset frequency threshold; when the performance decay rate of the join query is less than or equal to the preset decay threshold, the current index structure is maintained.

[0176] This embodiment provides an index maintenance mechanism for dynamic graphs; specifically, in the aforementioned scheme, the graph will continue to evolve due to context drift, the attachment of new nodes, and repeated retrievals.

[0177] If the index structure is not adjusted for a long time, although the scale of graph content nodes and topological edges continues to expand, query time may increase and hit rate may decrease. If the index is fully rebuilt every time, it will consume a lot of resources and affect online use. Therefore, this embodiment adopts the method of performance degradation monitoring + local index reconstruction, and only maintains the active subgraph.

[0178] The index maintenance module runs at a fixed period, such as once every 2 hours or per shift. At the end of each period, two metrics are calculated: one is the change in the average query time of the adaptive retrieval module in the current period; the other is the change in the retrieval hit rate. The changes here are usually compared with the previous period. The two changes are weighted and summed to obtain the joint query performance decay rate.

[0179] The weight coefficients used in the weighted summation are pre-allocated based on the impact of query time and retrieval hit rate on the overall availability of the system in the historical retrieval logs, and the sum of the weight coefficients corresponding to the change in query time and the change in retrieval hit rate is 1.

[0180] For ease of explanation, assuming the average query time in the previous period was 2 seconds and increased to 3 seconds in the current period, the change in time can be normalized to 0.5; the search hit rate in the previous period was 92% and decreased to 84% in the current period, the change in hit rate can be expressed as 0.08.

[0181] If the system sets the weights to 0.6 and 0.4 respectively, the performance degradation rate of the joint query can be approximated as follows: If the preset decay threshold is 0.25, then 0.332 exceeds the threshold, indicating that the current index structure is no longer suitable for the dynamic graph state and maintenance actions need to be triggered.

[0182] Next, the system does not perform a full index reconstruction of the entire reconstructed graph, but first counts the update frequency of each subgraph in the current cycle; the subgraphs here can be divided according to orders, processes, exception types or entity clusters; for example, the subgraph related to molding degumming has added 12 nodes and updated edge weights 26 times in the current cycle, which is a high update frequency.

[0183] The cutting tool wear subgraph is updated only once, which is considered a low update frequency. If the preset frequency threshold is 10 times per cycle, the former is considered an active subgraph, while the latter maintains its original index and is thus preserved.

[0184] Local index reconstruction can be achieved by reordering hot entities, updating the inverted index, refreshing the adjacency cache, or rebuilding the vector nearest neighbor index. However, the reconstruction scope is limited to active subgraphs. For example, in the degumming-related subgraph, the system prioritizes writing high-access entities such as humidity-abnormal glue batches, changes in viscosity sensors, and new nodes into the hot index area to shorten the query path for similar anomalies in subsequent queries.

[0185] The reason for introducing this mechanism is that the previous layer scheme can continuously expand the graph, but the query performance of the dynamic graph will degrade as the structure evolves; without index maintenance, the system will soon encounter the problem that the accuracy of the graph topology representation is improved, but the retrieval time complexity is significantly increased.

[0186] Furthermore, when the performance decay rate of the join query is less than or equal to a preset threshold, the system maintains the current index structure to avoid meaningless rebuilding; if the decay rate exceeds the threshold, but the current server resources are insufficient, the local rebuilding task can be split into multiple subtasks and executed in order of priority.

[0187] If the number of active subgraphs is too large, causing the local reconstruction range to approach the entire graph, the system can further increase the frequency threshold and prioritize the reconstruction of the top few subgraphs that are directly related to the current high-frequency anomaly type to prevent online service blockage.

[0188] During the two shifts following the aforementioned order anomaly, the engineer made multiple inquiries regarding the delamination of the molding material, while the system continuously uploaded new glue viscosity data to the relevant graphs.

[0189] As a result, the query time increased significantly and the hit rate decreased slightly in this cycle. The index maintenance module calculated that the performance decay rate of the joint query had exceeded the set threshold. Therefore, it only rebuilt the index for the active subgraph of the degumming anomaly, while maintaining the low update frequency subgraphs such as the cutting tool life. After completion, the response time of the same type of query in the next cycle returned to a low level.

[0190] The purpose of this step is to maintain the retrieval performance of the dynamic map without performing a full reconstruction, thereby achieving a balance between computational resource consumption and online query efficiency.

[0191] In a preferred embodiment of the present invention, the current batch material attribute characteristics include shoe upper material batch replacement records; workshop microclimate characteristics include workshop temperature and humidity fluctuation data; cross-modal causal relationship data include cross-modal characteristics that characterize molding delamination abnormalities caused by changes in glue curing time due to workshop temperature and humidity fluctuations.

[0192] This embodiment provides a specific expression mechanism for typical abnormal scenarios in shoe processing. Specifically, in the aforementioned context monitoring, material properties and microclimate have been used as context features. However, without explaining them in conjunction with typical failure mechanisms in actual shoe factories, it is still difficult to highlight why these features participate in causal chain retrieval.

[0193] Therefore, this embodiment uses the batch replacement record of shoe upper material and the temperature and humidity fluctuation data of the workshop as preferred features, and gives a cross-modal causal relationship organization method of temperature and humidity fluctuation - glue curing time change - molding delamination abnormality;

[0194] Batch replacement records for shoe upper materials can come from the material management system, work order switching logs, or manual entry records by the work team; these records not only indicate what materials were used, but more importantly, they mark whether unplanned replacements occurred.

[0195] This is because even if the material name is the same, different batches may have differences in surface treatment, adsorption, and softness, which will change the glue penetration and curing behavior; the system can encode no replacement as 0, replace as 1, and then add the batch difference vector before and after replacement as part of the material property characteristics;

[0196] The temperature and humidity fluctuation data in the workshop not only records the temperature and humidity values ​​at a certain moment, but also emphasizes the fluctuation process; for example, one hour before molding, the humidity rose from 58% to 72%, and the temperature rose from 24℃ to 27℃.

[0197] The system can extract four types of statistics: average value, maximum value, fluctuation range, and rate of change, forming a microclimate feature vector. For adhesive processes, the fluctuation range is often better able to explain the changes in curing time than a single point value.

[0198] In the graph, cross-modal causal relationship data can be organized as a chain of connections between different data modalities; taking this embodiment as an example, it includes at least three types of entities: first, environmental modal entities, such as 10:20 humidity rising to 72%;

[0199] Secondly, process or material modal entities, such as 09:10 replacing B27 glue batch and replacing the upper material of this batch; Thirdly, equipment or quality modal entities, such as 14:20 failure to compensate for the holding time of the molding press and 14:30 glue detachment in the quality inspection record.

[0200] When the system searches for mold delamination in the current context, it will prioritize searching for combined paths that can form a continuous explanatory chain. Assuming that the correlation degree between the environmental entity and the implicit node of glue curing time extension is 0.89, the correlation degree between the implicit node of material replacement and glue layer adhesion change is 0.83, and the correlation degree between curing time extension and delamination anomaly is 0.91, the system can output a cross-modal causal chain: workshop temperature and humidity fluctuation → glue curing time change → mold delamination anomaly, with batch replacement of shoe upper material as an enhancing factor.

[0201] The changes in adhesive curing time here can be obtained directly from viscosity sensor or process test data, or can be inferred from logs and time series segments; for example, if the adhesive layer stabilization time is significantly delayed in a high humidity environment under the same holding pressure time, this node can be used as an intermediate interpretation node in the spectrum.

[0202] Furthermore, when temperature and humidity fluctuations exist but no evidence related to adhesive curing time is observed, the system can still output environmental anomaly candidates, but set their confidence level to medium; when material replacement records exist and temperature and humidity are stable, the system is more inclined to output the path of material batch difference → adhesion anomaly.

[0203] If both exist but their interaction directions are inconsistent—for example, high humidity usually leads to slower curing, while current actual detection shows faster curing—the system will reduce the overall confidence of this path and indicate that other unobserved factors are involved.

[0204] In the aforementioned order, a batch of shoe upper materials were replaced in the morning due to inventory issues, and the humidity in the workshop continued to rise in the afternoon. When the system searched for the cause of the glue delamination, it not only hit the environmental data of increased humidity, but also the shoe upper material replacement record and glue batch change log. Combined with the viscosity data, it deduced that the glue curing time was prolonged.

[0205] The final output is not a single sensor alarm, but a change in the curing time of the glue caused by temperature and humidity fluctuations, which is amplified in the context of material replacement, and finally triggers a cross-modal feature chain characterized as delamination of the molded product.

[0206] The purpose of this step is to explicitly incorporate the most common and engineering-interpretive contextual factors in shoe manufacturing into the causal chain expression, thereby transforming the search results from a stack of related records to an understandable geometric consistency or mechanistic chain.

[0207] In a preferred embodiment of the present invention, the data acquisition module is deployed on an edge computing node; the graph database is deployed on a cloud server; and the adaptive retrieval module receives abnormal query commands and displays cross-modal causal relationship data through a human-computer interaction interface.

[0208] This embodiment provides a deployment mechanism for edge-cloud collaboration; specifically, in the aforementioned solution, the system functions are already clear, but if all calculations are concentrated in the same location, network latency, upload congestion, or interaction lag may easily occur on-site in the shoe factory.

[0209] In particular, since the device's time-series data is sampled at a high frequency, uploading all of it without preprocessing would increase the transmission burden; therefore, this embodiment adopts a collaborative structure of edge acquisition, cloud-based mapping, and terminal retrieval and display.

[0210] The data acquisition module is preferably deployed on edge computing nodes near each workshop or production line; the edge nodes can be directly connected to cutting machines, sewing machines, forming presses, temperature and humidity sensors, and field log terminals;

[0211] Its main tasks include real-time acquisition of raw data, completion of preliminary feature extraction and vectorization, execution of necessary time synchronization, and local caching of obvious anomalies;

[0212] The purpose of this setting is that high-frequency device data does not have to be uploaded in its entirety, but is first converted into a more compact physical representation before being sent to the cloud;

[0213] Graph databases are preferably deployed on cloud servers; the cloud has stronger storage and graph computing capabilities, making it suitable for maintaining complete order-level, batch-level, and process-level reconstructed graphs.

[0214] After the entity data uploaded by the edge nodes is aggregated in the cloud, it can be used together with historical data to participate in operations such as context baseline comparison, dynamic topology reconstruction, mounting of isolated data, and index maintenance; the cloud also facilitates the sharing of graph knowledge among multiple workshops and shifts, forming a more complete accumulation of anomaly experience;

[0215] The adaptive retrieval module can be deployed in the cloud or in a cloud-edge collaborative deployment, and provides query services to process engineers, quality inspectors or team leaders through a human-computer interaction interface;

[0216] The interface can be a workshop terminal, an industrial tablet, or a management center page. When a user enters an anomaly query command, the retrieval module reads cross-modal causal relationship data related to the order, process, or anomaly type from the current graph and displays it in the form of a visual path, a list of related entities, or a timeline.

[0217] A simplified data flow simulation is provided; at 09:00, the edge node of the cutting station collects a segment of equipment pressure time series data, which is then extracted locally and formed into a vector. ;

[0218] Humidity fluctuation data was collected at the edge node of the molding workshop at 10:20, forming a vector. The log terminal recorded the replacement of B27 glue at 09:10, forming a text vector. ; , , Upload each data point to a cloud-based graph database, and create nodes and edges in the cloud.

[0219] After the glue delamination occurred at 14:30, the engineer entered a query on the terminal. The cloud retrieval module returned the path of humidity fluctuation - glue batch change - delamination based on the current reconstructed map, and displayed it on the interface in chronological order.

[0220] The reason for adopting this deployment structure is that field data acquisition emphasizes real-time performance and resilience to network fluctuations, while map management and cross-batch retrieval emphasize centralized computing and historical accumulation; separating the two can balance stability and analytical depth.

[0221] Additionally, when the edge node experiences a brief disconnection from the cloud, the edge side continues to cache the vectorized results and the original timestamps, and retransmits them in batches after the network is restored; if cloud retrieval is temporarily unavailable, the human-computer interaction interface can first display the local alarms and recent time period summaries retained by the edge node.

[0222] If clock drift at an edge node causes timestamp inconsistencies, the system can perform secondary correction in the cloud using a unified time source before writing it into the official graph. If the same query requires data retrieval across workshops but some nodes have not uploaded in time, the interface should mark the result as a partial data view to prevent misleading operators.

[0223] In the aforementioned degumming anomaly, the cutting and forming workshops respectively collected data from local edge nodes and preprocessed it. All entities were uploaded to the cloud graph database to form a complete graph.

[0224] After an engineer enters an anomaly query on an industrial tablet in the molding workshop, the system returns a cross-modal causal chain from the cloud and displays the time, process, and correlation strength of each node on the interface. If the network is briefly interrupted at that time, the tablet first displays the humidity anomaly and glue change logs cached at the edge, and then completes the causal chain after the network is restored.

[0225] The purpose of this step is to achieve an engineering system architecture suitable for deployment in actual shoe factories by coordinating the local data collection on the edge side with the centralized map management on the cloud side, taking into account the real-time performance, data integrity and cross-modal joint retrieval capabilities.

[0226] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A shoe manufacturing process analysis system based on industrial big data, characterized in that, The system includes: The data acquisition module is configured to collect multi-source heterogeneous data from multiple process steps in the discrete manufacturing process of shoe processing, and to convert the multi-source heterogeneous data into feature vectors and unify them to the same dimension through linear projection or zero-padding operations, and then map them into data entities in a preset graph database. Based on the preset shoe processing process flow rules, the module establishes the initial association topology weights between the data entities. The context monitoring module is configured to extract production context features from the multi-source heterogeneous data in real time, obtain the deviation value between the production context features and the context baseline pre-constructed based on historical good product data by calculating the feature distance, and output the context drift status. The dynamic topology reconstruction module is configured to receive the context drift state. When the context drift state indicates that a context drift has occurred, it extracts the timestamp difference between data across process links as a spatiotemporal lag parameter and aligns the multi-source heterogeneous data. Based on the aligned multi-source heterogeneous data, it updates the association topology weights between data entities across process links in the graph database and generates a reconstructed graph. The adaptive retrieval module is configured to receive anomaly query instructions through a human-computer interaction interface, perform cross-modal joint retrieval based on the current production context features and the reconstructed graph, and extract and output cross-modal causal relationship data that matches the anomaly query instructions; The multi-source heterogeneous data includes equipment time-series data, environmental sensor data, and unstructured process logs.

2. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The context monitoring module includes: The feature extraction unit is configured to extract the current batch material attribute features and workshop microclimate features from the multi-source heterogeneous data as the production context features; The state determination unit is configured to: determine that context drift has occurred when the deviation value is greater than a preset drift threshold, and output a first context drift state; and output a second context drift state and maintain the current associated topology weight when the deviation value is less than or equal to the preset drift threshold.

3. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The dynamic topology reconstruction module includes: The spatiotemporal alignment unit is configured to align the device timing data and the unstructured process log based on the spatiotemporal lag parameter, and generate an aligned data sequence. The weight update unit is configured to update the associated topological weights using the aligned data sequence to generate the reconstructed graph; The cross-process steps include the cutting step, the sewing step, and the forming step.

4. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The adaptive retrieval module includes: The parsing unit is configured to parse the anomaly query instruction to obtain the target anomaly feature vector; The association calculation unit is configured to obtain the contextual association degree between the target anomaly feature vector and each of the data entities by calculating the similarity between the target anomaly feature vector and the feature vectors of each of the data entities in the reconstructed graph. The feature output unit is configured to: extract the corresponding data entity as the cross-modal causal relationship data output when the context relevance is greater than a preset relevance threshold; and add an irrelevant label to the corresponding data entity and filter it when the context relevance is less than or equal to the preset relevance threshold.

5. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The system also includes: The isolated data mounting module is configured to compare the data source network identifier of the input data stream with a preset data access whitelist to detect undefined detached data streams, extract the metadata features and spatiotemporal context features of the detached data streams, and vectorize and concatenate the metadata features and spatiotemporal context features. By calculating the similarity between the concatenated feature vector and the feature vectors of each data entity in the graph database, the matching probability between the detached data stream and each data entity in the reconstructed graph is obtained.

6. The shoe manufacturing process analysis system based on industrial big data according to claim 5, characterized in that, The isolated data mounting module is also configured to: When the matching probability is greater than the preset matching threshold, the free data stream is used as a new attribute graph node and attached to the data entity with the highest matching probability in the reconstructed graph. When the matching probability is less than or equal to the preset matching threshold, the free data stream is stored in the preset isolated storage area and a manual review prompt is generated.

7. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The system also includes: The index maintenance module is configured to periodically calculate the weighted sum of the change in query time and the change in retrieval hit rate of the adaptive retrieval module in the current period, as the joint query performance decay rate; When the performance decay rate of the joint query is greater than the preset decay threshold, a local index reconstruction instruction is triggered to reconstruct the index of active subgraphs in the reconstructed graph whose update frequency is greater than the preset frequency threshold, while maintaining the original index of subgraphs whose update frequency is less than or equal to the preset frequency threshold. When the performance decay rate of the joint query is less than or equal to the preset decay threshold, the current index structure is maintained.

8. The shoe manufacturing process analysis system based on industrial big data according to claim 2, characterized in that, The current batch material attribute characteristics include the shoe upper material batch replacement record; The workshop microclimate characteristics include workshop temperature and humidity fluctuation data; The cross-modal causal relationship data includes cross-modal characteristics that characterize abnormal debonding of the molding process due to changes in adhesive curing time caused by fluctuations in workshop temperature and humidity.

9. The shoe manufacturing process analysis system based on industrial big data according to claim 1, characterized in that, The data acquisition module is deployed on an edge computing node; The graph database is deployed on a cloud server; The adaptive retrieval module receives the abnormal query command through the human-computer interaction interface and displays the cross-modal causal relationship data.

Images