An industrial product data compression and traceability system and method

Abstract

This invention provides an industrial product data compression and traceability system and method, relating to the field of electronic digital data processing technology. The invention constructs a causal knowledge graph and calculates the criticality of causal paths at each node. Combined with the local information entropy of real-time data, it generates a "causal entropy weight" that comprehensively represents the overall value of data points. Subsequently, using this "causal entropy weight" as the sole driver, it adaptively determines refined compression and indexing strategies during the storage phase. Furthermore, during the traceability phase, it integrates with dynamic "causal proximity factors" to generate a priority extraction queue. This constructs a new data processing paradigm that deeply couples storage and traceability, guided entirely by causal knowledge.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to an industrial product data compression and traceability system and method. Background Technology

[0002] In modern industrial manufacturing, especially in high-precision fields such as semiconductors, new energy batteries, and precision instruments, production process data is characterized by massive volume (TB / PB levels), high dimensionality, and high frequency. To ensure product quality and achieve fault traceability, enterprises need to store this data long-term. Existing technologies typically employ general compression algorithms to reduce storage costs and perform a "decompression before retrieval" operation during traceability, resulting in low query efficiency. Furthermore, mainstream traceability methods rely on keyword matching or pre-defined encoding, making it difficult to discover deep, complex causal relationships across processes within massive amounts of data, creating a sharp contradiction between "data storage cost" and "traceability efficiency and depth." Therefore, maximizing the traceability value of data within limited storage costs is a crucial direction for ongoing exploration in the field of industrial big data.

[0003] Existing technologies typically face the following challenges when handling the compressed storage and quality traceability of industrial data:

[0004] The disconnect between storage strategies and data value: Existing data compression schemes often employ general algorithms, applying a uniform compression strategy to all data or based on simple statistical characteristics (such as rate of change). Such methods fail to identify the intrinsic value differences of different data points for potential future quality tracing. As a result, some "data" crucial for revealing causal chains may be over-compressed due to its apparent smoothness, leading to information loss, while a large amount of redundant "background noise" data consumes valuable storage resources. This storage approach results in a mismatch between storage costs and data value, creating a limitation of resource misallocation.

[0005] The "blindness" and superficiality of the traceability process: When conducting quality traceability, as disclosed in Chinese patent application CN117349306A, the method mainly relies on preset business association codes or keywords for retrieval. This type of method is effective when handling traceability tasks with clear identifiers. However, when faced with complex, cross-process, and non-obvious quality problems, this "strong label"-based retrieval method proves inadequate. It often returns a large amount of relevant but non-critical data, requiring secondary manual verification by engineers, resulting in a "data avalanche" phenomenon. More importantly, it struggles to reveal the hidden, deep-seated causal relationships behind the data, causing traceability to remain at the level of "finding relevant data" rather than delving into the level of "locating the root cause."

[0006] The "two-stage" separation of storage and traceability: In existing technological systems, data storage and data traceability are typically treated as two independent, sequentially executed technical stages. When data is stored, the system lacks consideration for its future traceability; conversely, during traceability, the query engine cannot utilize the valuable information potentially contained within the stored data to optimize the query path. This design separation leads to an inefficient "global decompression followed by brute-force retrieval" model, causing traceability response time to deteriorate dramatically with the increase in total data volume, creating an irreconcilable contradiction between "storage costs" and "traceability efficiency."

[0007] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide an industrial product data compression and traceability system and method to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for industrial product data compression and traceability, comprising the following steps:

[0011] S1: Construct a knowledge graph based on historical production data, which includes production entity nodes and causal edges representing the influence relationships between the production entity nodes. Calculate a preset causal path criticality for each production entity node based on the topology of the knowledge graph and the attributes of the causal edges.

[0012] S2: Receive the real-time industrial data stream corresponding to the production entity node online, and calculate the local information entropy that characterizes the local information uncertainty of the data point for the data point in the real-time industrial data stream;

[0013] S3: For the data point, generate a composite causal entropy weight based on the preset causal path criticality of its corresponding production entity node and its own local information entropy.

[0014] S4: Based on the value of the causal entropy weight, the causal entropy weight is mapped into a storage strategy vector containing storage control parameters through one or more preset continuous functions;

[0015] S5: Invoke the compression module and the indexing module, and use the storage control parameters contained in the storage strategy vector as the quality control input of the compression module and the density control input of the indexing module to perform parameterized compression storage and / or fine-grained index construction on the data points.

[0016] An industrial product data compression and traceability system, the system being used to execute the industrial product data compression and traceability method, comprising:

[0017] The graph construction module is used to construct a knowledge graph based on historical production data, which includes production entity nodes and causal edges representing the influence relationships between the production entity nodes. Based on the topology of the knowledge graph and the attributes of the causal edges, it calculates a preset causal path criticality for each production entity node.

[0018] Entropy generation module: used to receive real-time industrial data streams corresponding to the production entity nodes online, and to calculate local information entropy that characterizes the local information uncertainty of the data points in the real-time industrial data stream;

[0019] Causal entropy weight generation module: used to generate a composite causal entropy weight for the data point based on the preset causal path criticality of its corresponding production entity node and its own local information entropy;

[0020] Vector generation module: used to map the causal entropy weight into a storage strategy vector containing storage control parameters based on the value of the causal entropy weight through one or more preset continuous functions;

[0021] Execution Invocation Module: Used to invoke the compression module and the indexing module, using the storage control parameters contained in the storage strategy vector as the quality control input of the compression module and the density control input of the indexing module, so as to perform parameterized compression storage and / or fine-grained index construction on the data points.

[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: At the level of linkage between storage strategy and causal knowledge, this invention treats it as a value matching behavior guided in real-time by causal importance inferred from a knowledge graph; a core "causal entropy weight" calculation simultaneously determines "how data should be physically stored" and its "logical importance in the knowledge network." This fundamentally solves the aforementioned problem of the disconnect between storage strategy and data value, ensuring that physical storage resources always accurately serve the highest needs of logical traceability.

[0023] Secondly, at the level of linking tracing paths and data access, this invention transforms tracing queries from blind data retrieval into a precise navigation of underlying stored data directly guided by causal paths planned by a knowledge graph. Furthermore, this invention dynamically introduces a "causal proximity factor" during tracing, which is then fused with the inherent "causal entropy weight" during storage to generate a "tracing relevance score." This design allows static storage strategies and dynamic tracing needs to work synergistically, not only finding a logical explanation for the problem but also simultaneously generating the most efficient physical data access plan, thereby effectively overcoming the "blindness" and "two-stage" separation problems of the tracing process. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0025] Figure 2 This is a schematic diagram of the execution logic of step S1 of the present invention;

[0026] Figure 3 This is a schematic diagram illustrating the execution logic of steps S2 and S3 of the present invention;

[0027] Figure 4 This is a schematic diagram illustrating the execution logic of steps S4 and S5 of the present invention. Detailed Implementation

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0030] Example 1:

[0031] Please see Figures 1 to 4 The present invention provides a technical solution:

[0032] An industrial product data compression and traceability method, the execution entity of which is an industrial data processing system, which executes the method by its internal processor to execute corresponding computer program instructions; including the following steps:

[0033] S1: Construct a knowledge graph based on historical production data, which includes production entity nodes and causal edges representing the influence relationships between production entity nodes. Based on the topological structure of the knowledge graph and the attributes of the causal edges, calculate the preset causal path criticality for each production entity node.

[0034] S2: Receive real-time industrial data streams corresponding to production entity nodes online, and calculate the local information entropy that represents the uncertainty of local information for each data point in the real-time industrial data stream.

[0035] S3: For each data point, a composite causal entropy weight is generated based on the preset causal path criticality of its corresponding production entity node and its own local information entropy.

[0036] S4: Based on the value of the causal entropy weight, the causal entropy weight is mapped into a storage strategy vector containing storage control parameters through one or more preset continuous functions;

[0037] S5: Call the compression module and indexing module, and use the storage control parameters contained in the storage strategy vector as the quality control input of the compression module and the density control input of the indexing module to perform parameterized compressed storage and / or fine-grained index construction on the data points.

[0038] Further explanation: Construct a knowledge graph representing the production process; calculate the causal transit confidence for the causal edges of the knowledge graph; and, based on the causal transit confidence, obtain the causal path criticality of each production entity node through iterative calculation.

[0039] The steps for calculating the confidence level of causal transmission include: calculating it based on historical data using the Granger causality test;

[0040] The steps for calculating the confidence level of causal transmission further include: calculating the historical causal baseline value and the recent causal dynamic value respectively, and weighting and fusing the two through a time decay coefficient to obtain the time-weighted confidence level of causal transmission.

[0041] Further explanation: The step of obtaining the criticality of causal paths through iterative calculation employs an improved graph centrality algorithm;

[0042] The improved graph centrality algorithm is a variant of the web page ranking algorithm that uses time-weighted causal transit confidence as the transit weight of the edges and includes a damping coefficient.

[0043] Iterative calculations include initializing the criticality values ​​of causal paths for all production entity nodes and updating them cyclically until the change in the criticality value of the causal path for each production entity node is less than a preset convergence threshold.

[0044] The following is a detailed implementation description of the above content: The core technical feature of this embodiment is as follows: First, a knowledge graph representing the production process is constructed, and a "time-weighted causal transit confidence score" that integrates historical statistical patterns and recent data dynamics is calculated for the causal edges in the graph; then, this dynamic confidence score is used as a key input and applied to an improved graph centrality algorithm. Through iterative calculation, a "causal path criticality score" that accurately reflects the global importance of each production entity node in the knowledge graph in the entire dynamic causal network is determined. This method, by introducing a time decay mechanism, enables the evaluation of node importance to adaptively respond to changes in the production process, thereby significantly improving the accuracy and timeliness of criticality measurement.

[0045] The key parameters involved in this embodiment are defined as follows. All parameters are determined through subsequent calculation processes. After standardization, their values ​​are limited to the closed interval [0,1].

[0046] The time-weighted confidence score for causal transitivity has the following parameter notation: , is a dynamically calculated value at the current time t; logically, it represents a quantitative assessment of the conditional probability that a change in the state of upstream production entity node i during the production process will cause a corresponding change in the state of its directly downstream adjacent production entity node j. The acquisition and determination method is as follows: the calculation model originates from the application and improvement of the Granger Causality Test in statistics; by introducing a time decay mechanism, the results of the Granger Causality Test from different time windows are fused to obtain a confidence assessment that combines long-term stability and recent sensitivity. First, a historical causal baseline value is defined. This value is obtained by applying the Granger Causality Test to a pair of historical production data time series covering a relatively long time span (the past six months in this example), and is used to characterize the strength of a stable, long-term causal association between nodes. Second, a recent causal dynamic value is defined. This value is obtained by applying the same Granger Causality Test to a pair of data time series covering only the most recent shorter time span (the past week in this example), and is used to capture newly emerging or changing causal associations between nodes. The final time-weighted causal transit confidence score is obtained by weighted summation of historical causal baseline values ​​and recent causal dynamic values. The weights assigned to recent causal dynamic values ​​are determined by a time decay coefficient. The weights assigned to the historical causal baseline values ​​are determined by subtracting the time decay coefficient from "1". The value of the time decay coefficient is determined based on the characteristics of the production process. For rapidly changing processes, a larger value of 0.7 is used as an example, while for stable and mature processes, a smaller value of 0.2 is used as an example. To ensure the output value range, after weighted summation, the result is normalized to its maximum and minimum values, mapping it to the closed interval [0,1].

[0047] The higher the time-weighted confidence score of causal transit, the stronger and more significant the causal relationship from node i to node j is in recent times; conversely, the lower the score, the weaker the causal relationship is or it only exists in distant historical data.

[0048] In this embodiment, for the causal edge (i,j), the calculated historical causal baseline value is 0.6, and the recent causal dynamic value is 0.9. A time decay coefficient is set. The value is 0.3. The unnormalized confidence level is calculated as follows: 0.3 multiplied by 0.9, plus (1 minus 0.3) multiplied by 0.6, resulting in 0.69.

[0049] The criticality of the causal path for a production entity node, with parameter symbol . Logically, it represents the global importance or influence score of production entity node i within the causal network constituted by the entire knowledge graph. The higher the value, the more likely a small change in its state will have a significant impact on the quality of the final product through the transmission of the causal chain. The acquisition methods and determination procedures are as follows:

[0050] The computational model is an adaptive modification of the classic PageRank algorithm in graph theory. The core modification is to replace the uniform link propagation probability in the original algorithm with a time-weighted causal propagation confidence score with clear physical meaning, as defined in this embodiment.

[0051] The criticality of the causal path for production entity node i is composed of two parts. The first part is a basic importance value, calculated from a preset damping coefficient d1, specifically the difference between "1" and the damping coefficient. The second part is a cumulative contribution value, which is the sum of the importance contributions of all upstream neighboring production entity nodes j pointing to production entity node i. For each upstream production entity node j, its single contribution to production entity node i is calculated as follows: take the criticality of the causal path of production entity node j itself, divide it by the total out-degree of node j (i.e., the total number of causal edges originating from production entity node j), and then multiply it by the time-weighted causal transit confidence carried on the causal edges from production entity node j to production entity node i. Finally, this cumulative contribution value is multiplied by the damping coefficient d1. The damping coefficient d1 is a neighborhood standard value, set to 0.85 in this embodiment. The entire calculation process is performed through multiple iterations until the criticality values ​​of the causal paths of all production entity nodes converge to a stable state. To ensure the output value range, after the iteration converges, the final causal path criticality scores of all production entity nodes are normalized by the maximum and minimum values, and mapped to the closed interval [0,1].

[0052] The higher the criticality value of a causal path, the more important the production entity node is as the source or a necessary path in more causal chains, and the higher its criticality in production tracing.

[0053] In this embodiment, during one iteration, producer entity node i has only one upstream producer entity node j. Producer entity node j currently has a criticality of 0.5, a total out-degree of 2, a causal propagation confidence of edge (j,i) of 0.8, and a damping coefficient d1 of 0.85. Therefore, in this iteration, the contribution of producer entity node j to producer entity node i is: 0.5 divided by 2 and then multiplied by 0.8, resulting in 0.2. The updated causal path criticality of producer entity node i will include 0.85 multiplied by 0.2.

[0054] The complete computational process for achieving this core technical feature is broken down as follows:

[0055] 1.1) The initial input consists of two parts: First, structured data obtained from systems such as Manufacturing Execution System (MES) and Enterprise Resource Planning (ERP) that defines production entities (such as equipment, processes, and material batches) and their technological relationships; Second, historical time-series data associated with each production entity obtained from systems such as Supervisory Control and Data Acquisition (SCADA) and sensor logs.

[0056] 1.2) Processing Step 1: Based on the structured data, instantiate the production entities as production entity nodes in the knowledge graph. Based on the process flow relationships, establish directed causal edges between production entity nodes with upstream and downstream relationships, thus initially constructing the topology of the knowledge graph.

[0057] 1.3) Processing Step Two: Traverse each causal edge (i,j) in the knowledge graph. Extract the historical time series data corresponding to production entity node i and production entity node j. Calculate the time-weighted causal transit confidence of this causal edge. And use it as the weight attribute of that edge.

[0058] 1.4) Processing step three: Assign the same initial causal path criticality value to all production entity nodes in the knowledge graph. In this embodiment, it is set to the reciprocal of the total number of nodes.

[0059] 1.5) Processing Step Four: Enter the Iterative Loop. In each iteration, traverse all production entity nodes and use the causal path criticality values ​​obtained by other production entity nodes in the previous iteration to update the causal path criticality of the current production entity node.

[0060] Specifically, after each iteration, the change in the "causal path criticality" of all production entity nodes compared to the result of the previous iteration is calculated. If this change is less than a preset convergence threshold, the calculation is considered converged, and the iteration loop is exited. Otherwise, the next iteration continues; in this embodiment, the preset convergence threshold is set to one ten-thousandth.

[0061] 1.6) Processing step five: After the iteration converges, collect the final "causal path criticality" value of all production entity nodes; perform a global maximum and minimum normalization on all these values ​​to ensure that the final output value falls within the closed interval [0,1].

[0062] 1.7) The final output of the process is a mapping table or database record, which contains a unique identifier for each production entity node and its corresponding normalized causal path criticality. .

[0063] The core fusion mechanism of this embodiment is reflected in the calculation of causal path criticality, which is essentially an iterative propagation model based on a weighted directed graph. The innovative improvement of this embodiment lies in replacing the static, undifferentiated links in the propagation process with dynamic, time-weighted causal propagation confidence levels that contain rich causal information. This allows the fusion process to not only consider the network topology but also deeply integrate the inherent dynamic characteristics of the process data, thereby identifying "potential critical nodes" that are not structurally prominent but exhibit strong causality in recent dynamics.

[0064] Time decay coefficient The determination of this depends on domain knowledge. In a typical embodiment, it is jointly calibrated by a team of process engineers based on the process stability and product lifecycle of the specific production line. For a new product line that frequently adjusts process parameters, Set to 0.6-0.8 to quickly respond to changes in causal relationships caused by adjustments; for a stable production line that has been operating for many years, The value is set to 0.1-0.3 to focus on long-term, verified causal relationships.

[0065] The damping coefficient d1 is a standard hyperparameter of the graph centrality algorithm. Its function is to ensure the convergence of the iterative process and avoid the "sorting trap" problem. Based on a large number of published documents and industry practices, its value is determined to be 0.85 in this embodiment. It does not require frequent adjustments for specific scenarios, thus ensuring the universality and stability of the algorithm.

[0066] Time-weighted confidence level of causal transit With causal path criticality There is a clear positive correlation between them. The more high-confidence input edges a production entity node has, the greater its potential for increasing causal path criticality score.

[0067] Criticality of causal paths of upstream production entity nodes Criticality of causal path with downstream production entity nodes There is also a clear positive correlation between them. The importance of a production entity node is positively transmitted to its downstream production entity nodes through causal edges.

[0068] Further elaboration and demonstration of step S1 above:

[0069] The greater the criticality of the causal path, the higher the analysis priority of the corresponding production entity node during fault tracing; specifically:

[0070] When the criticality of the causal path The closer the value is to 1, the more critical the upstream source or core hub role that production entity node i plays in the entire production causal network; even small fluctuations in its own state will be amplified and transmitted to multiple downstream production stages through a series of high-confidence causal chains, ultimately having a greater impact on product quality. Such production entity nodes are given higher analysis priority when performing fault tracing.

[0071] When the criticality of the causal path The closer a value is to 0, the more likely the production entity node i is to be in a relatively isolated or terminal position in the causal network. Its state changes have a more limited impact on downstream links, or there are no significantly influential key nodes upstream. In retrospective analysis, such production entity nodes have a lower priority for analysis.

[0072] Criticality of the causal path affecting the final output The key input parameter is the time-weighted causal transit confidence score. .

[0073] When the graph topology remains unchanged, the time-weighted causal transit confidence score on any causal edge (j,i) pointing to the producing entity node i is... The increase will lead to a decrease in the criticality of the causal path of production entity node i. Monotonically increasing.

[0074] According to the computational model of the present invention, the causal path criticality of production entity node i It includes a cumulative contribution term, which is the sum of the contributions of all upstream production entity nodes j. The single contribution of each upstream production entity node j is correlated with the causal propagation confidence. Proportional. Therefore, the confidence level of causal transmission. The increase in directly and linearly increases the contribution value from the producing entity node j, thereby increasing the criticality of the causal path. This is added in the next iteration update. This positive correlation design accurately maps the physical logic of the real world: if a production entity node is affected by other nodes with higher "determinism" or "probability," then the "importance" transmission received by that production entity node should also be correspondingly enhanced. The design of this invention precisely quantifies this logical transmission process.

[0075] To verify the effectiveness of the present invention's method for quantifying the criticality of production entity nodes based on a dynamic causal relationship model, the following comparative experiment was designed. The experimental subjects were three critical production entity nodes in a simplified semiconductor etching process: node A1 (gas flow controller), node B1 (RF power source), and node C1 (chamber pressure sensor). Three typical production scenarios were set up to compare the differences and advantages of the present invention's method with two existing methods (Method 1: based solely on topology, i.e., without considering edge weights; Method 2: based solely on static causal relationships, i.e., using only historical data to calculate fixed weights) in evaluating node criticality. See Table 1 below for details.

[0076] Parameters / Scenario Scenario 1: Stable Production Phase Scenario 2: Process Parameter Adjustment Period Scenario 3: Early stage of equipment aging Input parameters - - - A1-B1 Historical Causal Baseline Values ​​(Experiment 1) 0.82 0.82 0.82 A1-B1 Historical Causal Baseline Values ​​(Experiment 2) 0.81 0.81 0.81 A1-B1 Recent Causal Dynamic Values ​​(Experiment 1) 0.8 0.95 0.83 A1-B1 Recent Causal Dynamic Values ​​(Experiment 2) 0.83 0.94 0.81 B1-C1 Historical Causal Baseline (Experiment 1) 0.75 0.75 0.75 B1-C1 Historical Causal Baseline (Experiment 2) 0.76 0.76 0.76 B1-C1 Recent Causal Dynamic Values ​​(Experiment 1) 0.74 0.76 0.55 B1-C1 Recent Causal Dynamic Values ​​(Experiment 2) 0.77 0.74 0.58 Time decay coefficient 0.2 0.7 0.5 Intermediate parameters (Inventive Method - Experiment 1) - - - A1-B1-Confidence of Causal Transmission 0.816 0.911 0.825 B1-C1-Confidence of Causal Transmission 0.748 0.757 0.65 Output parameters: Criticality of causal path - - - Method 1: Based solely on topology (Experiment 1) 0.618 0.618 0.618 Method 1: Based solely on topology (Experiment 2) 0.618 0.618 0.618 Method 2: Based on static causal relationships (Experiment 1) 0.715 0.715 0.715 Method 2: Based on static causal relationships (Experiment 2) 0.712 0.712 0.712 Inventive Method (Experiment 1) 0.713 0.802 0.721 Inventive Method (Experiment 2) 0.718 0.795 0.715

[0077] 1. Compare Scenario 1 with existing technologies to verify the effectiveness of the basic approach:

[0078] In Scenario 1 (stable production phase), recent data is highly consistent with historical patterns (e.g., the historical causal baseline value is 0.82, and the recent value is 0.80). The causal path criticality of node B1 calculated by the method of this invention (average 0.716) is very close to the result of Method 2 (based on static causal relationships, average 0.714), and both are significantly higher than Method 1 (based solely on topology, 0.618). This proves that when the production state is stable, the evaluation results of this invention can be consistent with static methods based on long-term statistical patterns, verifying the fundamental rationality and feasibility of the algorithm design.

[0079] 2. In contrast to Scenario 2, highlighting the sensitivity to "dynamic changes" is the core point of this embodiment: In Scenario 2 (process parameter adjustment period), the engineer optimized the parameters of the gas flow controller (node ​​A1), which significantly enhanced its impact on the RF power source (node ​​B1), as reflected in the recent causal dynamic value of the causal transfer confidence soaring to 0.95.

[0080] Since Method 1 and Method 2 do not consider or cannot capture this recent dynamic, their output criticality values ​​are unchanged compared to Scenario 1 (both are 0.618 and 0.715), completely ignoring this important process change.

[0081] The method of this invention significantly improves the criticality of the causal path of node B1 to 0.802 (Experiment 1) by introducing a time decay coefficient (set to 0.7 in this scenario, focusing on recent changes). This represents an improvement of 12.2% compared to Method 2.

[0082] This data powerfully demonstrates the significant technical effectiveness of the core innovation of this invention—the time-weighted fusion mechanism. This mechanism enables criticality assessment to respond in real time and sensitively to dynamic changes in the production process, accurately identifying nodes whose criticality has been temporarily enhanced due to process adjustments.

[0083] 3. Compared with Scenario 3, highlighting the ability to identify "causal weakening": In Scenario 3 (early stage of equipment aging), the RF power source (node ​​B1) begins to experience performance degradation, which leads to a decrease in its ability to control the downstream chamber pressure (node ​​C1), as reflected in the recent causal dynamic value of the causal transfer confidence level significantly decreasing to 0.55.

[0084] Method 1 and Method 2 are also unable to perceive this change, and their output criticality values ​​remain unchanged.

[0085] The method of this invention captures the weakening of the causal chain B1→C1. Although the causal relationship A1→B1 remains stable, after comprehensive system evaluation, the criticality of the causal path of node B1 slightly increases to 0.721 compared to the stable period (Scenario 1) (Experiment 1). This may seem counterintuitive, but the underlying logic is that although the certainty of B1→C1 decreases, the "uncertainty" of B1 as a critical node increases. From a tracing perspective, this "out-of-control" node is more worthy of attention. More importantly, if the tracing target is node C1, the system will reduce the weight of tracing back from B1 due to the decrease in the confidence of causal transmission.

[0086] This dataset demonstrates another significant advancement of the invention: it can not only identify causal reinforcement, but also quantitatively identify the weakening or failure of causal relationships.

[0087] Further explanation: The steps for calculating local information entropy include: using a sliding time window to acquire the data sequence of the real-time industrial data stream, discretizing and binning the data sequence, and calculating the local information entropy based on the Shannon entropy formula;

[0088] Discretized binning employs adaptive binning boundaries determined based on historical data percentiles;

[0089] The steps for calculating local information entropy also include: dividing the calculated original entropy value by the theoretical maximum entropy value based on the number of bins to obtain the normalized local information entropy;

[0090] The step of generating composite causal entropy weights is to use a weighted summation of the criticality of the causal path and the normalized local information entropy.

[0091] The criticality weighting factor and entropy weighting factor used in the weighted summation are dynamically determined based on the global process fluctuation degree that characterizes the overall state of the system, calculated in real time.

[0092] The following are specific implementation instructions for the above content:

[0093] The core technical feature of this embodiment lies in the following: For real-time industrial data streams received online, an entropy calculation method based on adaptive binning of historical data distribution is used to quantify local information entropy that accurately reflects the uncertainty of local information in the data. Subsequently, this entropy value is fused with a preset normalized causal path criticality that characterizes the global importance of the data source. This fusion process employs an adaptive weighting mechanism based on production status to generate a final causal entropy weight that dynamically reflects the comprehensive traceability value of data points. Through adaptive binning and adaptive weighting, the accuracy and robustness of data value assessment under different operating conditions are significantly improved.

[0094] The key parameters involved in this embodiment are defined as follows. All parameters are determined through subsequent calculation processes. After standardization, the output value range is limited to the closed interval [0,1].

[0095] Normalized local information entropy, whose parameter symbol is: This is a dimensionless numerical value. Logically, it represents the relative level of signal uncertainty or information content in the real-time industrial data stream within the current sliding time window. The acquisition method and determination mechanism are as follows:

[0096] The computational model originates from the classic Shannon-Entropy formula in information theory, with innovative improvements to its input data preprocessing. Specifically, it employs an adaptive binning method based on historical data percentiles, replacing traditional equal-width or equal-frequency binning to eliminate the impact of data offset and scaling variations on the entropy calculation results. The computational logic is as follows: First, offline analysis of the historical data of the target production entity nodes is performed to calculate multiple key percentile points of their numerical distribution; in this example, these are the 10th, 20th, ..., 90th percentiles. These percentile points will serve as the adaptive binning boundaries for subsequent online calculations. During online processing, the data sequence within the current sliding time window is acquired, and each data point in the sequence is binned using the pre-calculated adaptive binning boundaries. Subsequently, the frequency of data points appearing in each bin is counted and substituted into the Shannon-Entropy calculation paradigm, i.e., the probability of occurrence of each bin is multiplied by its base-2 logarithm, and then the results of all binning are summed and the negative is taken to obtain an original local entropy. To achieve normalization, the original entropy value is divided by a theoretical maximum entropy value, which is the entropy value when the data is uniformly distributed across all bins. This maximum entropy value is calculated as the logarithm of the total number of bins to the base 2. The final quotient is the normalized local information entropy.

[0097] The larger the value of the normalized local information entropy, the more drastic the fluctuations and the more irregular the pattern of the current data stream, and the more potential abnormal information it contains; the smaller the value, the more stable and predictable the data stream is.

[0098] In this embodiment, offline analysis identified 10 bins (9 boundary points). If the data points within the current sliding window are evenly distributed across these 10 bins, the calculated value of the original local entropy is the logarithm of 10 to the base 2, approximately 3.32. At this point, the theoretical maximum entropy is also 3.32, therefore the final normalized local entropy is 1.0. If all data points fall into the same bin, the original entropy is 0, and the normalized local entropy is also 0.

[0099] Causal entropy weight, its parameter symbol is , is a dimensionless numerical value. Logically, it represents the comprehensive traceability value of a data point in the current production state, reflecting the macroscopic importance of its source (determined by the criticality of the causal path). (reflected) and its own microscopic information content (derived from normalized local information entropy) The result of the integration (reflected).

[0100] The innovation of the causal entropy weight calculation model lies in the fact that its weight coefficients are not fixed, but dynamically determined by the current global production state; the calculation logic is based on the criticality of the causal path between the production entity nodes corresponding to the data points. and the local information entropy normalized to the data point itself The weighted sum is then performed to obtain the result. Specifically, it is the normalized causal path criticality multiplied by the criticality weight factor. The normalized local information entropy is multiplied by an entropy weighting factor. Then add the two products together. The sum of the keyness weighting factor and the entropy weighting factor is always one.

[0101] The larger the causal entropy weight, the higher the comprehensive traceability value of that data point, and the more storage and indexing resources should be allocated to it. This embodiment is currently in "new process trial" mode, with a criticality weight factor... Set to 0.3, entropy weighting factor Set to 0.7. Data points are derived from the criticality of causal paths. A node with a local entropy of 0.9 has a local information entropy of 0.9. The value is 0.4. Therefore, its causal entropy weight is... The calculation is: 0.3 multiplied by 0.9, plus 0.7 multiplied by 0.4, the result is 0.55.

[0102] The complete computational process for achieving this core technical feature is broken down as follows:

[0103] 2.1) The initial input consists of two parts: first, a pre-calculated criticality of all production entity nodes and their normalized causal paths. The first is the mapping table; the second is the real-time industrial data stream received online from a specific production entity node.

[0104] 2.2) Processing Step 1: Maintain a fixed-length sliding time window on the real-time industrial data stream;

[0105] At each time step, obtain the complete data sequence contained within the window;

[0106] 2.3) Processing Step Two: Discretize and bin the data sequence within the window using a preset adaptive binning boundary. Calculate the normalized local information entropy of the current time window. .

[0107] 2.4) Step 3: Query the production status judgment module to obtain the current production mode.

[0108] Based on the retrieved production mode, select the corresponding key weight factor from the preset configuration table. and entropy weighting factor .

[0109] 2.5) Processing Step Four: Based on the source of the data flow, retrieve the normalized causal path criticality of the corresponding production entity node from the input mapping table. Using the weighting factors determined in step three, the criticality of the causal path is integrated. and local information entropy The final causal entropy weight is calculated. .

[0110] 2.6) The final output is the causal entropy weight calculated for the current time window. This value will serve as the direct basis for subsequent data compression and indexing strategy selection.

[0111] The innovative improvement of this embodiment lies in "adaptability," that is, the weighting coefficient. and Instead of being globally fixed hyperparameters, they are dynamically adjusted based on higher-dimensional contextual information—production status. This design enables simple linear models to simulate complex nonlinear decision-making logic, automatically switching their focus under different operating conditions. This significantly improves the model's adaptability and accuracy in changing industrial environments while maintaining its simplicity.

[0112] Causal path criticality With causal entropy weight There is a positive correlation between them. Normalized local information entropy. With causal entropy weight There is a positive correlation between them.

[0113] The following is a further explanation of the implementation of the above content:

[0114] When causal entropy weight The closer a value is to 1, the higher the retrospective value of that data point. This stems from two possible or overlapping scenarios: first, the data point itself exhibits dramatic, non-stationary fluctuations; second, the data point originates from a crucial node in a macroscopic causal network. Regardless of the scenario, a high... Each value signifies that the data point is crucial evidence for diagnosing potential problems and tracing causal chains, and therefore must be allocated the highest level of storage and indexing resources.

[0115] When causal entropy weight The closer a value is to 0, the lower its traceability value. This indicates that the data point originates from a non-critical node in a causal network, and its data performance is very stable and predictable. This type of data belongs to the "background noise" in the normal production process and is stored using a high compression ratio strategy to save costs, without the need to build a detailed index.

[0116] In another embodiment: the step of determining the global process volatility includes:

[0117] Obtain the normalized causal path criticality and real-time normalized local information entropy of each of the multiple pre-set production entity nodes within the production system.

[0118] For each of the multiple production entity nodes, calculate the product of the criticality of the causal path and the local information entropy; and

[0119] The global process variability is obtained by summing the products calculated from all production entity nodes.

[0120] The steps for dynamically determining the criticality weighting factor and the entropy weighting factor include:

[0121] Based on the global process volatility, an entropy weighting factor is calculated, where the value of the entropy weighting factor is positively correlated with the value of the global process volatility; and

[0122] Based on the entropy weighting factor and a pre-defined rule that the sum of the two weighting factors is a constant, the criticality weighting factor is calculated.

[0123] The steps for calculating the entropy weighting factor further include:

[0124] Multiplying the global process volatility by a preset dynamic response adjustment factor yields an intermediate product; and

[0125] The intermediate product is compared with the preset upper limit of the weight, and the smaller of the two values ​​is taken as the entropy weight factor.

[0126] Set the global process volatility, with the parameter symbol as follows: is a dimensionless numerical value that is updated in real time during each calculation cycle and has a range within the closed interval [0,1]. Logically, it represents the degree of macroscopic uncertainty or disorder of the entire production system at the current moment. The calculation model is a weighted average model, its innovation lying in traversing all production entity nodes in the knowledge graph. For each node, its normalized causal path criticality is calculated. Instead of the local information entropy normalized at the current moment Multiply the products at all nodes. Then, sum the products obtained at all nodes; the final sum is the global process volatility. This is due to the criticality of causal paths. and local information entropy The value range is [0,1], and in practical applications, the criticality of the causal path of all nodes is... The sum is usually not 1, to ensure The value range strictly falls within [0,1], and it is necessary to consider the values ​​of all nodes. Perform a renormalization to make the sum equal to 1, and then perform a weighted summation.

[0127] The closer the global process volatility value is to 1, the more severe the data fluctuations are occurring simultaneously at multiple, especially critical, production nodes in the system, indicating that the system as a whole is in a state of "turbulence" or "abnormal exploration." The closer the value is to 0, the more likely the entire system is to be in a highly stable and predictable state of "smooth operation."

[0128] In this embodiment, if there are only two production entity nodes in the system, production entity node 1... It is 0.8. The value is 0.1; for production entity node 2 It is 0.2. The value is 0.3. Since the sum is 1, the criticality of the causal path remains unchanged after renormalization. Therefore, the global process volatility... The calculation is: 0.8 multiplied by 0.1, plus 0.2 multiplied by 0.3, the result is 0.14.

[0129] Keyness weighting factor With entropy weighting factor : Through a deterministic function, derived from the global process volatility It is directly calculated. This calculation model is a nonlinear mapping control law; when the system is stable, we should place more trust in prior causal knowledge based on long-term data; when the system is turbulent, we should pay more attention to real-time entropy increase signals that indicate the source of chaos. The specific determination logic is as follows:

[0130] A configurable dynamic response adjustment factor is introduced, with parameter notation as follows: The dynamic response adjustment factor takes a value greater than or equal to 1 and is used to adjust the sensitivity of weight changes to system fluctuations. The larger the value, the greater the volatility of the weighting factor with the global process. The faster the rate of change.

[0131] For entropy weighting factor The calculation method is as follows: global process fluctuation. With dynamic response adjustment factor Multiply the products and then take the smaller value between the product and 1.0; to ensure the entropy weighting factor. The value will not exceed 1.

[0132] Keyness weighting factor The calculation method is as follows: subtract the calculated entropy weighting factor from 1.0. .

[0133] With global process variability The entropy weighting factor increases continuously from 0 to 1. The criticality weighting factor will then increase non-linearly and continuously. The number will decrease accordingly and continuously.

[0134] The above weight adjustments are made smoothly and continuously as the system state changes, improving the model's granularity and responsiveness to minute state changes. The system no longer relies on external, potentially lagging production plan information to determine its own state. It develops an intrinsic understanding of the global state by calculating the entropy and criticality of all nodes in real time, and uses this understanding to guide its behavior, forming a complete and advanced adaptive feedback control loop.

[0135] Further explanation: Based on the causal entropy weight that characterizes the value of data points, the storage strategy is determined; and upon receiving a tracing request, relevant data is extracted according to a preset causal relationship.

[0136] Storage control parameters include: fidelity parameters for controlling the fidelity of the compression algorithm, and index granularity parameters for controlling the index building density;

[0137] The steps for extracting relevant data include: dynamically calculating the tracing relevance score for data points along the causal tracing path; and determining the extraction priority of data points based on the tracing relevance score.

[0138] The calculation of the relevance score is based on the causal entropy weight corresponding to the storage of fused data points, and the causal proximity factor that characterizes the causal distance between the data points and the defective nodes.

[0139] The causal proximity factor is calculated based on the number of path hops between the production entity node where the data point is located and the defect node on the causal tracing path, and is determined by an exponential decay function.

[0140] The steps to determine the extraction priority of data points are as follows: sort all relevant data points in descending order based on the relevance score to generate a priority data extraction queue.

[0141] To further explain the above: The core technical feature of this embodiment is that it constructs a deeply coupled closed-loop system from data storage to intelligent traceability, which generates significant synergistic gains;

[0142] During the data storage phase, this method assigns a causal entropy weight to each data point. Through a nonlinear mapping function, it is continuously and in real time transformed into a function containing a fidelity parameter. and index granularity parameters Storage strategy vector .

[0143] During the data tracing phase, this method introduces a tracing relevance score. The score is dynamically calculated for each data point along the causal tracing path after a tracing request is received. A weighted model is used to incorporate the inherent causal entropy weights of the data points at storage. And the causal proximity factor characterizing its spatiotemporal proximity to the current defect event The data is then integrated. Finally, based on this relevance score, the system generates a priority data extraction queue, prioritizing the extraction and decompression of the data with the highest score.

[0144] This design establishes a dynamic information feedback bridge between the storage and traceability stages, enabling static storage strategies and dynamic traceability requirements to work synergistically.

[0145] The key parameters involved in this embodiment are defined as follows. All parameters are determined through subsequent calculation processes. After standardization, the output value range is limited to the closed interval [0,1].

[0146] Storage strategy vector, whose parameter notation is is a two-dimensional vector. Logically, it encapsulates the complete storage operation instructions to be executed for a single data point; its specific content is determined by the causal entropy weight of the data point. Uniquely determined. This vector contains the following two components:

[0147] Fidelity parameter, its symbol is , is a dimensionless numerical value with a closed range of [0,1]. Logically, it represents the level of information fidelity that should be preserved when performing lossy compression on data points. The calculation model for this parameter originates from the logistic function in mathematics, which smoothly maps the input value to the (0,1) interval. This invention utilizes the characteristics of this function to weight causal entropy. The linear change is transformed into a nonlinear, saturated change of the fidelity parameter. The calculation method involves converting the input causal entropy weight... The curve is centered (i.e., the center point offset is subtracted; in this embodiment, the center point offset is set to 0.5), and then multiplied by a fidelity sensitivity factor used to control the steepness of the curve. This result is then used as the negative power of the natural exponent, and after calculating the exponent value, "1" is added. Finally, the reciprocal is taken.

[0148] The smaller the causal entropy weight, the smaller the fidelity parameter, and the lower the corresponding fidelity, resulting in a higher compression ratio; specifically:

[0149] When causal entropy weight When it approaches 0, the fidelity parameter Approaching 0 corresponds to low fidelity, at which point the compression ratio is higher; when the causal entropy weight... When it approaches 1, A value close to 1 corresponds to high fidelity, approaching lossless compression.

[0150] In this embodiment, the center point offset is 0.5, and the fidelity sensitivity factor is 10. If the input causal entropy weight... If the value is 0.8, then after centering, multiplying by the sensitivity factor yields 3. The calculation result is approximately 1 divided by (1+e⁻³), which gives... The value is 0.953. This value will be passed to the compression algorithm as its quality control parameter.

[0151] The index granularity parameter, its parameter symbol is: is a dimensionless numerical value with a closed range of [0,1]. Logically, it represents the density or granularity of the index built for that data point. The calculation model for this parameter originates from a power function. Its design principle is that the consumption of index resources should have a superlinear relationship with the growth of value, meaning that disproportionately more index resources should be invested in high-value data. The calculation method involves using the input causal entropy weight... Using the preset index enhancement index as the base, the index enhancement index is used as the exponent and a power operation is performed. The result is the index granularity parameter. In this embodiment, the index enhancement index is greater than 1.

[0152] When causal entropy weight The lower the index granularity parameter, the better. The lower the value, the more pronounced the causal entropy weight. With the increase of index granularity parameters The value of accelerates. When the causal entropy weight When it is 1, the index granularity parameter It is also 1.

[0153] In this embodiment, the index enhancement index is 2.0. If the input causal entropy weight... If it is 0.8, then the index granularity parameter The calculated value is 0.8 squared, which is 0.64. If the causal entropy weight... If it is 0.2, then the index granularity parameter It is only 0.04.

[0154] The relevance score is traced back to its parameter symbol. is a dimensionless numerical value with a closed range of [0,1]. Logically, it represents the comprehensive importance of a data point relative to a specific traceability request, integrating its static value and dynamic correlation. Its calculation depends on the following parameters:

[0155] The causal proximity factor, whose parameter symbol is: , is a dimensionless numerical value with a closed range of [0,1]. Logically, it quantifies the "causal distance" between a node on a causal tracing path and the defect originating node. The calculation model originates from the concept of distance decay in graph theory, specifically employing an exponential decay function. The calculation logic is as follows: on the identified causal tracing path, calculate the shortest path hop count between the target node and the defect originating node. Then, multiply this shortest path hop count by a negative proximity decay coefficient, and use the calculation result as the exponent of the natural constant e, performing a power operation. The result is the causal proximity factor. In this embodiment, the proximity decay coefficient is greater than 0, used to control the decay rate.

[0156] The shortest path hop count of the defective node itself is 0. The value is 1. The farther a node is from the defective node, the larger its shortest path hop count. The closer the value is to 0, the better. In this embodiment, the proximity attenuation coefficient is 0.5. If a node on the path is 3 hops away from the defect node, then... It is calculated as e raised to the power of (-0.5×3), and is approximately 0.223.

[0157] The complete computational process for realizing this core technical feature can be broken down into the following two independent but interrelated stages:

[0158] Phase 1, Adaptive Storage Process:

[0159] The initial input is the causal entropy weights that have already been calculated for the data points. ; Input-based causal entropy weights Calculate the fidelity parameters respectively. and index granularity parameters The calculated and Combined into a storage strategy vector Call the compression module, passing in the data points and... The compression operation is performed using the compression quality parameters. The index module is called, passing in the metadata of the data points. The index module is based on The value of determines the specific index behavior; if If the value is greater than 0.9, then create a separate entry for that point; if... If the value is less than 0.1, only the statistical summary information of the corresponding data block is updated. The final output is the compressed data block and the updated index structure.

[0160] Phase Two: Causal Tracing and Priority Extraction Process

[0161] The initial input is a traceability request containing information about defects in the target product;

[0162] The request is parsed to locate the defective node in the knowledge graph, and then propagated backward along the causal edges to identify one or more high-probability causal tracing paths.

[0163] Iterate through each node in the causal tracing path, and for each stored data point associated with these nodes within the relevant time frame, perform the following sub-steps:

[0164] Calculate the causal proximity factor of the production entity node containing the data point relative to the defective node. .

[0165] Read the causal entropy weight assigned to the data point at storage from the stored metadata. .

[0166] Fusion and Calculate the traceability relevance score for this data point. .

[0167] Calculate all The data points are put into a set and then sorted according to... The values ​​are sorted in descending order to generate a priority data extraction queue.

[0168] The final output is a priority data extraction queue; subsequent data extraction and decompression operations will be performed strictly in the order of this queue, thereby ensuring that the most relevant information is presented to the user first.

[0169] During the tracing phase, a tracing relevance score is calculated. The core is the integration of causal entropy weights and causal proximity factor This embodiment employs a linear weighted average method. This method was chosen because it effectively balances information from two different dimensions while ensuring interpretability. This represents the static importance of the data points, while This represents its dynamic relevance in this specific tracing task. The ingenuity of this invention lies in using dynamically calculated causal proximity factors that are strongly correlated with the query itself. It is incorporated into the prioritization decision-making process for data extraction.

[0170] Relevance score The calculation model is as follows: the tracing relevance score equals the static importance weight multiplied by the causal entropy weight, plus the dynamic relevance weight multiplied by the causal proximity factor. The sum of these two weight coefficients is 1. Specific values ​​are calibrated through offline simulation experiments, and the optimization objective is to minimize the "average number of decompressions required to find the true root cause data." Since in specific tracing tasks, the direct causal relationship with the defect is often more critical than its generalized historical importance, the dynamic relevance weight in this embodiment is assigned a relatively high value between 0.6 and 0.8.

[0171] To further elaborate on the above: In this embodiment, the core output parameter during the traceability phase is the traceability relevance score. Its range is limited to the closed interval [0,1].

[0172] Relevance score The larger the value, the higher the reference value of the data point for the current tracing task;

[0173] When tracing relevance scores The closer the value is to 1, the more the characterization system judges the data point as the "prime suspect" evidence for the current specific tracing task. This is the result of two high-value attributes combined: first, the data point is judged to have high static and universal value during storage; second, it is highly proximate to the source of the current defect event in the causal chain. The value indicates that the data point should be given the highest extraction priority and should be immediately decompressed and presented to the analysts.

[0174] When tracing relevance scores The closer a value is to 0, the lower its reference value for the current tracing task. This stems from its low static value, its significant distance from the defect event in the causal chain, or both. The value implies that the data point should be placed at the end of the priority queue, and the system should avoid wasting I / O and computing resources on such data.

[0175] The final output traceability score is affected. The key input parameter is the causal entropy weight. and causal proximity factor .

[0176] When other parameters remain constant, causal entropy weight The increase will lead to a higher relevance score. Monotonically increasing. According to the weighted fusion model of the present invention, The calculation formula includes a term that is the static importance weight multiplied by Since this weight is a non-negative constant, therefore and A linear positive correlation exists between them. This design maps to the following physical logic: data points that have been repeatedly proven to be key and information-rich in history have causal entropy weights. The higher the "prior probability" of it as potential evidence in any retrospective investigation, the higher the probability.

[0177] When other parameters remain constant, the causal proximity factor The increase will also lead to a decrease in the relevance score. Monotonically increasing. Similarly, The calculation formula includes another term, which is the dynamic correlation weight multiplied by Since this weight is also a non-negative constant, a linear positive correlation exists between the two. The rationale behind this design is that events closer to the point where the problem occurs are more likely to be the direct or indirect cause of the problem. This invention... The parameters quantify and integrate this universal law into the sorting model, ensuring the logical focus of the tracing.

[0178] To verify the significant technical effect of the priority tracing method based on biphasic dynamic weights of this invention, the following comparative experiment was designed. The experimental objective was to evaluate the efficiency of different methods in locating the true root cause data for a defect tracing request of "battery bulging". The comparative methods included Method 1: tracing based on static value (based solely on causal entropy weight sorting at storage time) and Method 2: causal path tracing without optimization (extracting all data on the causal path, without internal priority). The key evaluation index of the experiment was the "average extraction rank of root cause data". The lower the rank, the higher the efficiency of the method; see Table 2 below for details:

[0179] Parameters / Scenario Scenario 1: High-value, close-range root causes Scenario 2: Low-value, close-range root causes Scenario 3: High-value, long-distance interference Input parameters (root cause data points) - - - Causal entropy weight (Experiment 1) 0.92 0.15 0.95 Causal Entropy Weight (Experiment 2) 0.91 0.16 0.94 Causal proximity factor (Experiment 1) 0.95 0.98 0.12 Causal proximity factor (Experiment 2) 0.96 0.97 0.11 Output parameters: Relevance score (inventive method) - - - Relevance scoring for retrospective analysis (Experiment 1) 0.938 0.648 0.452 Relevance scoring (Experiment 2) 0.94 0.646 0.442 Output parameter: Average rank of root cause data extraction - - - Method 2: Unoptimized Traceability (Experiment 1) 52 48 55 Method 2: Unoptimized Traceability (Experiment 2) 49 53 51 Method 1: Based on static value (Experiment 1) 1 58 2 Method 1: Based on static value (Experiment 2) 1 55 1 Inventive Method (Experiment 1) 1 1 45 Inventive Method (Experiment 2) 1 1 48

[0180] 1. Compare with scenario two to verify the ability to discover the root causes of "hiddenness":

[0181] In scenario two, the real root cause is its low static value (causal entropy weight) (average 0.155), but high causal proximity to the defect (average 0.975).

[0182] Method 1 (based on static value) was completely misled by noisy data with high static value, ranking the true root cause at 56.5 (average), resulting in low tracing efficiency and almost equivalent to failure.

[0183] This embodiment's method, through a dynamic fusion mechanism, despite its low causal entropy weight, achieves a high causal proximity factor, boosting its relevance score to 0.647 (average), standing out among all relevant data and ultimately ranking first in extraction. This dataset powerfully demonstrates the decisive advantage of this invention in fusing dynamic relevance. Compared to Method 1, which relies solely on static value, the invented method improves the discovery ranking of root cause data by over 98%, solving the major technical challenge of existing technologies' inability to effectively identify "occasional and hidden" problems. This proves the feasibility and significant value of this invention.

[0184] Compare with scenario three to verify the ability to suppress "misleading" interference:

[0185] In scenario three, there is a highly interfering data point on the causal path with a high static value (average 0.945) but a great distance from the current defect (average causal proximity factor 0.115).

[0186] Method 1 again misjudged the data, ranking this irrelevant and distracting data in 1.5th place (on average), which misled the analysts and wasted valuable analysis time.

[0187] In this embodiment, the method effectively suppresses the final traceability relevance score to 0.447 (average) due to the low causal proximity factor of the data point, reducing its final extraction ranking to 46.5 (average), thus successfully avoiding interference. This dataset demonstrates the significant effect of this invention in improving the signal-to-noise ratio. By introducing causal proximity as a "filter," this invention can effectively suppress "noise" that, while important in itself, is irrelevant to the current problem, ensuring the focus of the traceability process. Its anti-interference capability far exceeds that of existing technologies.

[0188] In all scenarios, Method 2 (without optimization) performed the worst, with an average extraction rank of approximately 51.3. Method 1 was unstable, performing perfectly in Scenario 1 but failing completely in Scenarios 2 and 3, with an average rank of 20.0. The inventive method demonstrated extremely high robustness in all scenarios, with an average extraction rank of 15.7.

[0189] Experimental data irrefutably demonstrates that this invention, through its innovative biphasic dynamic weight fusion mechanism, achieves a breakthrough improvement in traceability efficiency, accuracy, and robustness compared to existing technologies. It not only quickly identifies problems but also accurately pinpoints the "right" problems, realizing a qualitative leap from "data extraction" to "intelligent diagnosis."

[0190] Based on a retrospective analysis of hundreds of historical real-world fault tracing cases, statistical modeling was performed on the distribution of tracing relevance scores for root cause data, strongly correlated evidence, and weakly correlated evidence. Incorporating the experience of senior process diagnostic experts, the following SMART criteria were developed to optimally match data extraction operations with the cognitive load of analysts; see Table 3 below for details.

[0191] Priority range Tracing Relevance Score - Interval Technical Operations and System Behavior Primary root cause candidate region [0.8,1.0] Automatic and real-time operation: The system automatically marks all data points within this interval as "Level 1 root cause candidates". It immediately initiates the highest priority I / O request to perform full-fidelity decompression and highlights the decompressed data sequence in the "Core Evidence" panel of the traceability analysis interface. Secondary supporting data area (0.5,0.8) Prefetching and on-demand loading: The system automatically marks data points within this range as "secondary supporting evidence data." A medium-priority I / O request is initiated to prefetch compressed data into the cache, but decompression is not performed immediately. Summary information of these data points is listed in the "Related Evidence" panel of the interface; real-time decompression and visualization are triggered only after the user clicks on it. Background context area [0,0.5] Index-only, manual query: The system does not perform any active data operations. Index entries for these data points are only displayed in the "Background Data" list on the interface. Analysts must manually initiate explicit queries to trigger the extraction and decompression of this data to prevent information overload.

[0192] The "causal entropy weight" designed in this embodiment serves as the core technical action, achieving deep linkage control at two levels and generating a significant synergistic gain effect:

[0193] The linkage between storage strategy and causal knowledge: The system's data compression and indexing strategies are no longer static or based solely on the statistical characteristics of the data itself, but are guided in real-time and dynamically by the causal importance inferred from the knowledge graph. A core "causal entropy weight" calculation simultaneously determines "how data should be physically stored" and its "logical importance in the knowledge network." This linkage ensures that physical storage resources always precisely serve the highest demands of logical tracing, achieving a dual optimization of storage efficiency and tracing efficiency.

[0194] Linkage between tracing paths and data access: The system's tracing queries no longer blindly search through the entire dataset, but are directly guided by causal paths planned by the knowledge graph, forming precise access to the underlying stored data. This core linkage seamlessly connects macroscopic causal reasoning with microscopic data extraction. A core graph path search action not only finds the logical explanation of the problem but also generates the most efficient data physical access plan. This embodiment transforms the tracing process from a computationally intensive data mining task into a lightweight, knowledge-driven navigation task, achieving a dual breakthrough in tracing depth and response speed.

[0195] Example 2:

[0196] Please see Figure 1 An industrial product data compression and traceability system, wherein the system is used to execute an industrial product data compression and traceability method, comprising:

[0197] The knowledge graph construction module is used to construct a knowledge graph based on historical production data, which includes production entity nodes and causal edges representing the influence relationships between production entity nodes. Based on the topology of the knowledge graph and the attributes of the causal edges, it calculates the preset causal path criticality for each production entity node.

[0198] Entropy generation module: used to receive real-time industrial data streams corresponding to production entity nodes online, and to calculate local information entropy that represents the uncertainty of local information of data points in the real-time industrial data stream;

[0199] Causal entropy weight generation module: used to generate composite causal entropy weights for data points based on the preset causal path criticality of their corresponding production entity nodes and their own local information entropy.

[0200] Vector generation module: Based on the value of causal entropy weight, it maps the causal entropy weight into a storage strategy vector containing storage control parameters through one or more preset continuous functions;

[0201] The execution call module is used to call the compression module and the indexing module. It takes the storage control parameters contained in the storage strategy vector as the quality control input of the compression module and the density control input of the indexing module to perform parameterized compressed storage and / or fine-grained index construction on the data points.

[0202] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0203] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An industrial product data compression and traceability method, characterized by, The specific steps include: S1: constructing a knowledge graph including production entity nodes and causal edges representing the influence relationship between the production entity nodes based on historical production data, and calculating a preset causal path criticality for each production entity node according to the topological structure of the knowledge graph and the attributes of the causal edges; S2: online receiving real-time industrial data stream corresponding to the production entity nodes, and calculating local information entropy representing local information uncertainty of a data point in the real-time industrial data stream; S3: generating a composite causal entropy weight for the data point according to the preset causal path criticality of the corresponding production entity node and the local information entropy of the data point itself; The step of calculating the local information entropy includes: obtaining a data sequence of the real-time industrial data stream using a sliding time window, discretizing and binning the data sequence, and calculating the local information entropy according to a Shannon entropy formula; The discretization and binning uses adaptive binning boundaries determined based on historical data percentiles; The step of calculating the local information entropy further includes: dividing the calculated original entropy value by a theoretical maximum entropy value based on the number of bins to obtain a normalized local information entropy; The step of generating the composite causal entropy weight uses a weighted sum of the causal path criticality and the normalized local information entropy; The criticality weight factor and the entropy weight factor used in the weighted sum are dynamically determined based on a global process fluctuation degree representing the overall state of the system; S4: based on the value of the causal entropy weight, mapping the causal entropy weight to a storage strategy vector containing storage control parameters through a preset continuous function; and after receiving a traceability request, extracting relevant data according to a preset causal relationship; The step of extracting relevant data includes: dynamically calculating a traceability correlation score for data points on the causal traceability path; and determining the extraction priority of the data points according to the traceability correlation score; The calculation of the traceability correlation score is based on the fusion of the causal entropy weight corresponding to the storage of the data point and a causal proximity factor representing the causal distance of the data point from the defect node; The calculation of the causal proximity factor is based on the path hop count of the production entity node where the data point is located and the defect node on the causal traceability path, and is specifically determined by an exponential decay function; S5: calling a compression module and an index module, using the storage control parameters contained in the storage strategy vector as the quality control input of the compression module and the density control input of the index module, to perform parameterized compression storage and / or refined index construction on the data points.

2. The method of claim 1, wherein: construct a knowledge graph representing a production process; calculate a causal path criticality for each production entity node in the knowledge graph based on the causal transmission confidence of the causal edges in the knowledge graph; The step of calculating the causal transmission confidence includes: based on historical data, calculating by Granger causality test; Specifically, calculate a historical causal benchmark value and a recent causal dynamic value respectively, and weight and fuse them through a time decay coefficient to obtain a time-series weighted causal transmission confidence. ​ 3. The method of claim 2, wherein: The step of iteratively calculating the causal path criticality degree uses an improved graph centrality algorithm; The improved graph centrality algorithm is a variant of a web page ranking algorithm, which uses the time-weighted causal transmission confidence as the transmission weight of an edge and includes a damping coefficient; The iterative calculation includes initializing the causal path criticality degrees of all production entity nodes and cyclically updating until the variation of the causal path criticality degree of each production entity node is less than a preset convergence threshold; The greater the causal path criticality degree, the higher the analysis priority of the corresponding production entity node in fault backtracking.

4. The industrial product data compression and backtracking method according to claim 3, characterized in that: The greater the normalized local information entropy, the more intense the fluctuation of the current data stream, the more irregular the pattern, and the more potential abnormal information contained; The greater the causal entropy weight, the higher the comprehensive backtracking value of the data point, and the more storage and index resources allocated to it.

5. The method of claim 4, wherein: The step of determining the global process fluctuation degree includes: obtaining the normalized causal path criticality degrees of a plurality of preset production entity nodes in the production system and real-time normalized local information entropy; for each production entity node, calculating the product of the causal path criticality degree and the local information entropy; and accumulating and summing the products calculated for all production entity nodes to obtain the global process fluctuation degree; The step of dynamically determining the criticality weight factor and the entropy weight factor includes: based on the global process fluctuation degree, calculating the entropy weight factor, wherein the value of the entropy weight factor is positively correlated with the value of the global process fluctuation degree; and according to the entropy weight factor and based on a preset rule that constrains the sum of the two weight factors to be a constant, calculating the criticality weight factor; The step of calculating the entropy weight factor further includes: multiplying the global process fluctuation degree by a preset dynamic response adjustment factor to obtain an intermediate product; and comparing the intermediate product with a preset weight upper limit value and taking the smaller one as the entropy weight factor.

6. The method of claim 5, wherein: Based on the causal entropy weight representing the value of the data point, the storage strategy is determined; The storage control parameters include a fidelity parameter for controlling the fidelity of the compression algorithm and an index granularity parameter for controlling the index construction density.

7. The industrial product data compression and backtracking method according to claim 6, characterized in that: The step of determining the extraction priority of the data point is specifically: descendingly sorting all relevant data points according to the backtracking correlation score to generate a priority data extraction queue.

8. The method of claim 7, wherein: The smaller the causal entropy weight, the smaller the fidelity parameter, the lower the corresponding fidelity, and the higher the compression ratio at this time; The greater the backtracking correlation score, the higher the reference value of the data point for the current backtracking task.

9. An industrial product data compression and traceability system characterized by: The system is used to execute the industrial product data compression and backtracking method according to any one of claims 1-8, and includes: a graph construction module configured to construct a knowledge graph comprising production entity nodes and causal edges representing influence relationships between the production entity nodes based on historical production data, and to calculate a preset causal path criticality for each of the production entity nodes according to a topology of the knowledge graph and attributes of the causal edges; an entropy generation module configured to receive online real-time industrial data streams corresponding to the production entity nodes, and to calculate, for a data point in the real-time industrial data streams, a local information entropy representing local information uncertainty of the data point; a causal entropy weight generation module configured to generate, for the data point, a composite causal entropy weight according to the preset causal path criticality of a production entity node corresponding to the data point and the local information entropy of the data point itself; a vector generation module configured to map the causal entropy weight to a storage strategy vector comprising storage control parameters by a preset continuous function based on a value of the causal entropy weight; an execution calling module configured to call a compression module and an index module, and to use storage control parameters comprised in the storage strategy vector as quality control input of the compression module and density control input of the index module, so as to perform parameterized compression storage and / or refined index construction on the data point.

Images