Information Management System for Digital Machining Workshop

Abstract

This invention relates to the field of digital manufacturing and industrial information management technology, and discloses an information management system for a digital processing workshop, comprising: acquiring equipment abnormal event records to generate a subgraph of equipment fault propagation with observational evidence; calculating process similarity to screen source domain products; fusing source domain prior causal graphs to perform Bayesian causal structure learning to generate a posterior cross-process quality causal subgraph; calculating conditional Granger causal relationships to generate a set of cross-layer causal edges; assembling a two-layer causal graph; establishing an extended structured Bayesian network model; performing structured variational inference to output joint attribution results; and performing propagation path tracing to generate an attribution report with uncertainty quantification. This invention solves the technical problems in related technologies of digital processing workshop fault attribution, such as the inability to effectively integrate multi-source heterogeneous information from the equipment layer and process layer, the difficulty in completing cross-process causal structure learning under limited sample conditions, and the lack of uncertainty quantification in the attribution results.

Description

Technical Field

[0001] This invention relates to the field of digital manufacturing and industrial information management technology, and more specifically, to an information management system for a digital manufacturing workshop. Background Technology

[0002] In digital manufacturing workshops, tracing the root causes of final product quality anomalies faces the challenge of two coupled mechanisms: equipment failure and process quality transmission. Multiple processing machines share infrastructure such as hydraulic stations and power networks, and the propagation effects of multiple independent failure sources superimpose on downstream equipment, forming mixed anomaly characteristics. Simultaneously, processing deviations in upstream processes are transmitted step-by-step through the process chain and may amplify to the point of exceeding final quality tolerances. Furthermore, the sample size is severely insufficient in the early stages of new product trial production.

[0003] In existing technologies, equipment failure tracing methods trace the root cause of failures by performing reverse traversal on the equipment event propagation graph and calculating the path propagation credibility, while process quality tracing methods construct causal relationship graphs between processes and perform quality tracing by performing conditional independence tests and causal effect estimations on all process data. The two methods operate independently.

[0004] The aforementioned existing technologies have the following technical problems: the equipment-level reverse traversal method implicitly assumes a single fault source, and when the propagation effects of multiple independent fault sources are superimposed, it cannot distinguish the superposition of multiple causes, thus omitting the real fault source; the process-level causal discovery method suffers from severe distortion of the causal graph structure when the sample size is insufficient in the early stage of new product trial production; the two methods of equipment fault tracing and process quality tracing operate independently, resulting in attribution conflicts and omissions in the cross-influence area, making it impossible to conduct cross-level joint tracing of product quality anomalies. Summary of the Invention

[0005] This invention provides an information management system for digital processing workshops, which solves the technical problems in related technologies such as the inability to effectively integrate multi-source heterogeneous information from the equipment layer and the process layer during the fault attribution process of digital processing workshops, the difficulty in completing cross-process causal structure learning under a small number of sample conditions, and the lack of uncertainty quantification in the attribution results.

[0006] This invention discloses an information management method for a digital processing workshop, comprising: acquiring abnormal event records of all equipment within the current fault time window; clustering abnormal event execution times to generate composite abnormal event clusters; extracting a backward reachability subgraph on the equipment event propagation graph and using the composite abnormal event clusters as node observation evidence to generate an equipment fault propagation subgraph; acquiring the process routes of the current product and historical products; calculating a process similarity matrix based on a weighted combination of process type matching degree, processing method consistency, key parameter range overlap rate, and processing material family similarity to screen source domain products; extracting cross-process causal relationship graphs of each source domain product and overlaying and fusing them to generate a fused prior causal graph skeleton; using prior confidence as the prior distribution of edge existence probability; performing Bayesian causal structure learning in conjunction with a small amount of sample data of the current product to generate a posterior cross-process quality causal subgraph; establishing equipment-process mapping relationships, and for each pair... The Granger causality test statistic for equipment nodes and process nodes is calculated to generate a set of cross-layer causal edges containing the effects of time delay and intensity. The equipment fault propagation subgraph, the posterior cross-process quality causal subgraph, and the set of cross-layer causal edges are assembled into a two-layer causal graph. An extended structured Bayesian network model containing latent variables of fault sources, propagation and reception, process self-deviation, inter-process transmission and reception, and cross-layer transmission are established on the two-layer causal graph. Structured variational inference is performed to obtain the approximate posterior distribution of all latent variables, and the probability ranking of independent equipment fault sources, the contribution ranking of process self-deviation, and the contribution ranking of cross-layer transmission paths are extracted. For root causes whose posterior probability of fault sources exceeds the judgment threshold, propagation path tracing is performed, and the posterior estimates of the intensity of each edge effect are multiplied along the path and the confidence interval of the cumulative contribution is calculated to generate a multi-source, multi-layer joint attribution report with uncertainty quantification.

[0007] Furthermore, the step of performing time-based clustering of abnormal events to generate composite abnormal event clusters includes: for an abnormal event sequence sorted by timestamp on the same device, calculating the time interval between adjacent events sequentially; if the time interval is less than the typical event response time of the device, then the event is grouped into the same composite abnormal event cluster as the previous event; otherwise, a new composite abnormal event cluster is created; before performing time-based clustering, semantic normalization processing is performed on the original abnormal event records to uniformly map fault descriptions with different expressions but the same meaning to predefined semantic event type codes.

[0008] Further, the weighted combination calculation of the process similarity matrix based on process type matching degree, processing method consistency, key parameter range overlap rate, and processing material family similarity includes: for the i-th process of the current product and the j-th process of a historical product, calculating the process similarity as a weighted sum of the process type matching degree, processing method consistency, key parameter range overlap rate, and processing material family similarity multiplied by their respective weight coefficients, where the sum of each weight coefficient is 1; the process type matching degree is a binary index, taking a value of 1 when the two processes belong to the same process type and 0 otherwise; the processing method consistency is a binary index, taking a value of 1 when the two processes use the same processing method and 0 otherwise. The value is 0; the overlap rate of the key parameter range is obtained by averaging the ratio of the intersection length to the union length of the specification range intervals of each key parameter shared by the two processes; the similarity of the processed material families is obtained by normalizing the ratio of the shortest path length of the leaf nodes corresponding to the two materials in the material family classification hierarchy tree to the maximum depth of the hierarchy tree; based on the process similarity matrix, for each process of the current product, the process with the highest process similarity in each historical product is selected to form the optimal matching pair, the process similarity values ​​of all optimal matching pairs are summed and divided by the total number of processes of the current product to obtain the normalized process similarity total score, and the top M historical products with the highest process similarity total score are selected as source domain products.

[0009] Furthermore, the step of extracting cross-process causal relationship graphs of each source domain product and overlaying and fusing them to generate a fused prior causal graph skeleton includes: for any potential causal edge in the fused prior causal graph skeleton, the occurrence of the potential causal edge in the causal graphs of the M source domain products is weighted and averaged with the total process similarity score between each source domain product and the current product to obtain the prior confidence of the potential causal edge; if there are causal edges with opposite directions between the same pair of process nodes in different source domain products, the causal edges with conflicting directions are marked as a state to be verified, and the posterior existence probability of the two directions of the causal edge of the state to be verified is calculated respectively during the Bayesian causal structure learning process, and the direction with the higher posterior probability is selected as the final causal edge direction.

[0010] Furthermore, the generation of the posterior cross-process quality causal subgraph further includes: extracting the workpiece-level full-process feature summary vector from the workpiece data of the current product that has completed all processes; preprocessing the measured values ​​of key quality parameters of each process using Z-score standardization; using the prior confidence as the prior distribution of the edge existence probability, and combining the likelihood function of a small number of samples to calculate the posterior existence probability of each edge; deleting edges with a posterior existence probability lower than a threshold, and retaining edges with a probability higher than a threshold; for each retained causal edge, using the effect intensity parameter of the corresponding edge in the source domain product as the prior mean, updating the posterior distribution of the effect intensity of each retained edge through Bayesian linear regression, and obtaining the posterior mean and posterior variance of the effect intensity of each retained edge; wherein, the Bayesian linear regression uses the standardized value of the quality parameter of the parent node process of the causal edge as the independent variable and the standardized value of the quality parameter of the child node process as the dependent variable, and solves the posterior distribution of the effect intensity by maximizing the posterior probability under Gaussian prior constraints.

[0011] Further, the calculation of the conditional Granger causality test statistic for each pair of equipment nodes and process nodes includes: establishing a regularized vector autoregressive model for the time series of equipment abnormal events and the time series of process quality parameter deviations, wherein the regularized vector autoregressive model uses the process quality parameter deviation as the dependent variable and the lagged terms of the process quality parameter deviation, the lagged terms of the equipment abnormal event time series, and the lagged terms of the condition variable as independent variables, wherein the condition variable is the quality parameter time series of all parent node processes of the process node in the posterior cross-process quality causality subgraph; applying L1 regularization constraints to the influence coefficient vector of equipment abnormal events on process quality parameters, and performing a significance test on the non-zero components in the influence coefficient vector after regularization solution to determine whether the cross-level causal relationship is valid; the influence delay is determined by selecting the lag order that maximizes the Granger causality test statistic, and the effect strength is estimated by introducing the variance increment of the model after introducing the equipment abnormal variable.

[0012] Furthermore, in the extended structured Bayesian network model: the conditional probability distribution of the observed variables of each node in the equipment layer is determined by linearly superimposing the effects of the fault source latent variable and each propagation and reception latent variable, and then mapping them to the independent occurrence probability of each semantic event type through a logistic function; the observed variables of each node in the process layer follow a Gaussian distribution with the mean being the product of the process's own deviation latent variable, the inter-process transmission and reception latent variable and the corresponding effect intensity coefficient, and the cross-layer transmission latent variable, wherein the prior distribution of the effect intensity coefficient is determined by the posterior mean and posterior variance of the effect intensity output by Bayesian causal structure learning; the conditional probability parameters of the equipment layer nodes and the process layer nodes are both trained through maximum a posteriori estimation, with the sum of the log likelihood of the observed variables and the prior log probability of the parameters as the optimization objective.

[0013] Further, the step of performing structured variational inference to obtain the approximate posterior distribution of all latent variables includes: defining a family of variational distributions to approximate the joint posterior distribution of all latent variables, wherein the family of variational distributions adopts a structured mean-field approximation, and the latent variables are divided into equipment-level latent variable groups, process-level latent variable groups, and cross-level latent variable groups according to the hierarchical structure of the two-layer causal graph. The variational distribution is decomposed into the product of the three groups of variational distributions, the dependencies between latent variables are preserved within the group, and the relationships between groups are processed independently; the optimization objective of variational inference is to maximize the lower bound of evidence, wherein the lower bound of evidence includes a reconstruction expectation term that measures the degree to which the observed data is explained by the model and a KL divergence regularization term that measures the difference between the variational posterior distribution and the prior distribution of latent variables; in each iteration, the other two groups of variational parameters are fixed in sequence, and gradient ascent updates are performed on the variational parameters of the current group. The three groups of alternating updates constitute a complete round of coordinate ascent iteration, and the process is repeated until the change in the lower bound of evidence between two adjacent iterations is lower than a preset convergence threshold.

[0014] Further, the step of multiplying the posterior estimates of the strength of each side effect along the path and calculating the confidence interval of the cumulative contribution includes: for each tracking path, obtaining the posterior mean of the strength of each side effect sequentially along the path and multiplying them to obtain the cumulative strength of the tracking path; normalizing the cumulative strength of all tracking paths to obtain the contribution ratio of each tracking path to the final quality deviation; based on the assumption that the posterior distributions of each side effect strength are independent, the approximate variance of the cumulative strength of the tracking path is equal to the square of the product of the posterior means of each side effect multiplied by the sum of the relative variances of each side effect strength, thereby obtaining the confidence interval of the cumulative contribution of the tracking path; as the number of trial production workpieces of new products increases, the data of the newly added workpieces are taken as new observations, and Bayesian incremental updates are performed on the basis of the current posterior distribution. The current posterior distribution is taken as the prior distribution of the new round of updates and multiplied with the likelihood function of the newly added data to obtain the updated posterior distribution. When the posterior existence probability of the causal graph structure changes significantly, the reassembly of the two-layer causal graph and the parameter update and variational inference of the extended structured Bayesian network model are triggered.

[0015] This invention discloses an information management system for a digital processing workshop, comprising: a subgraph generation module for generating equipment fault propagation subgraphs, used to acquire abnormal event records of all equipment within the current fault time window, cluster the execution time of abnormal events to generate composite abnormal event clusters, extract a backward reachability subgraph on the equipment event propagation graph and use the composite abnormal event clusters as node observation evidence to generate an equipment fault propagation subgraph; a source domain product screening module, used to acquire the process routes of current products and historical products, calculate the process similarity matrix based on a weighted combination of process type matching degree, processing method consistency, key parameter range overlap rate and processing material family similarity, and screen source domain products; a posterior causal subgraph generation module, used to extract cross-process causal relationship graphs of each source domain product, overlay and fuse them to generate a fused prior causal graph skeleton, use prior confidence as the prior distribution of edge existence probability, combine a small amount of sample data of the current product to perform Bayesian causal structure learning, and generate a posterior cross-process quality causal subgraph; and a cross-layer causal edge generation module, used to establish equipment-process mapping relationships. The system employs a multi-layer causal graph assembly and modeling module. This module calculates the conditional Granger causality test statistic for each pair of equipment nodes and process nodes, generating a set of cross-layer causal edges that includes the effects of time delay and intensity. A two-layer causal graph assembly and modeling module assembles the equipment fault propagation subgraph, the posterior cross-process quality causal subgraph, and the set of cross-layer causal edges into a two-layer causal graph. An extended structured Bayesian network model containing latent variables of fault sources, propagation and reception, process-specific deviations, inter-process transmission and reception, and cross-layer transmission is established on this graph. A variational inference and attribution extraction module performs structured variational inference to obtain the approximate posterior distribution of all latent variables, extracting the probability ranking of independent equipment fault sources, the contribution ranking of process-specific deviations, and the contribution ranking of cross-layer transmission paths. An attribution report generation module performs propagation path tracing for root causes whose posterior probability of the fault source exceeds a judgment threshold. It multiplies the posterior estimates of the effect intensity along the path and calculates the confidence interval of the cumulative contribution, generating a multi-source, multi-layer joint attribution report with uncertainty quantification.

[0016] The beneficial effects of this invention are as follows:

[0017] This invention addresses the technical problem of missing true fault sources during single-path backward traversal when the propagation effects of multiple independent fault sources are superimposed by introducing latent variables of fault sources and propagation and reception at the equipment layer and performing joint posterior inference. It achieves the technical effect of decomposing mixed anomaly representations into components contributing from different fault sources. Furthermore, by fusing prior causal graphs of historical products with similar processes and performing causal structure learning within a Bayesian framework using limited sample data, it solves the technical problem of severe distortion of the causal graph structure due to insufficient sample size in the early stages of new product trial production. It achieves the technical effect of maintaining the basic rationality of the causal graph structure and outputting attribution conclusions with uncertain intervals under small sample conditions. Finally, by unifying the modeling of equipment fault propagation mechanisms and process quality transmission mechanisms through a two-layer causal graph structure and introducing cross-layer transmission latent variables to perform joint inference, it solves the technical problem of attribution conflicts and omissions arising from the independent operation of two types of tracing methods in the cross-influence region. It achieves the technical effect of distinguishing the components of process quality parameter deviations from equipment faults in the current process, transmission from upstream processes, and the processing conditions of the current process itself. Attached Figure Description

[0018] Figure 1 This is a flowchart of the information management method for a digital processing workshop according to the present invention;

[0019] Figure 2 This is a timeline distribution diagram of equipment malfunction events according to the present invention;

[0020] Figure 3 This is a comparison chart of the similarity of historical product processes of this invention;

[0021] Figure 4 This is a comparison diagram of the prior confidence and posterior existence probability of causal edge in this invention;

[0022] Figure 5 This is a graph showing the intensity and significance of cross-layer causal edge effects in this invention;

[0023] Figure 6 This is a diagram illustrating the variational inference iterative convergence process of this invention;

[0024] Figure 7 This is the posterior probability / posterior norm diagram of the equipment layer and process layer attribution in this invention;

[0025] Figure 8 This is a graph showing the multi-source, multi-layer joint attribution contribution ratio and confidence interval of the present invention. Detailed Implementation

[0026] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0027] This embodiment provides an information management method for a digital processing workshop, such as... Figure 1 As shown, it includes the following steps:

[0028] Step 1: Obtain equipment abnormal event records and generate a subgraph of equipment fault propagation with observational evidence;

[0029] The system retrieves abnormal event records for all devices within the current fault time window from the log database of the workshop information management system. These records are then structured, parsed, and semantically normalized before being grouped by device identifier. For each device, time-based clustering is performed on the abnormal events. Consecutive events with time intervals shorter than the typical event response time for that device are aggregated into composite abnormal event clusters. Each composite abnormal event cluster contains multiple semantic event types and a time range. All backward reachable subgraphs terminating at the current alarm device are extracted from the device event propagation graph. The composite abnormal event clusters of each device node are used as node observation evidence to generate a device fault propagation subgraph with observation evidence.

[0030] It should be noted that the above time clustering adopts a sequential scan clustering method based on a time interval threshold. The input is a sequence of abnormal events ordered by timestamp on the same device and a typical event response time threshold for the device. The output is the clustering result that divides continuous events into several composite abnormal event clusters. Specifically, the time interval between adjacent events is calculated sequentially. If the interval is less than the preset typical event response time, the event is grouped into the same composite abnormal event cluster as the previous event; otherwise, a new composite abnormal event cluster is created. The typical event response time can be obtained from the statistics of historical abnormal events of this type of device, for example, by taking the median of the time intervals of historical continuous abnormal events of the device.

[0031] It should be noted that the above semantic normalization refers to uniformly mapping different expressions of fault descriptions with the same meaning in the original abnormal event records to predefined semantic event type codes. For example, low hydraulic system pressure and hydraulic pump pressure drop are uniformly classified into the same semantic event type.

[0032] Step 2: Obtain historical product process routes, calculate process similarity, and filter source domain products;

[0033] The process route definition and process parameter specifications for each step of the current product are retrieved from the process database. Simultaneously, the process routes for all products with sufficient accumulated processing data are searched in the historical product database. A step-by-step comparison is performed between the current product's process route and the process routes of each historical product. A process similarity matrix is ​​calculated based on a weighted combination of factors including process type matching degree, processing method consistency, overlap rate of key parameter ranges, and similarity of processed material families.

[0034] For the current product's first The first process and the first historical product Each process, process similarity The calculation formula is:

[0035] ;

[0036] ;

[0037] in, This serves as an index for the current product's process. An index for historical product processes; This represents the process type matching degree, with a value of 1 when two processes belong to the same process type and 0 otherwise. To ensure consistency in processing methods, the value is 1 when two processes use the same processing method, and 0 otherwise. The key parameter range overlap rate has a value range of [value range missing]. This indicates the overlap ratio of the specification ranges of key process parameters between two processes; To determine the similarity of the processed material families, the value range is: Calculated based on the hierarchical distance between the two materials in the material family classification system; , , , The weight coefficients for each dimension satisfy... The values ​​for the above four dimensions have all been normalized to [value missing]. For intervals with consistent dimensions, weighted summation operations can be performed directly.

[0038] Based on the process similarity matrix, the total process similarity score between each historical product and the current product is calculated, and the products with the highest total process similarity scores are selected. One historical product serves as the source domain product, among which The preset number of source domain products.

[0039] It should be noted that the above key parameter ranges overlap. The calculation method is as follows: For each key parameter shared by the two processes, calculate the ratio of the intersection length to the union length of their specification range intervals, and then take the average of the ratios for all shared key parameters. If the two processes do not share any key parameters, then the key parameter range overlap rate is... The value is 0.

[0040] It should be noted that the above-mentioned families of processed materials are similar in nature. The calculation method is as follows: Assume the material family classification system is a hierarchical tree, where each leaf node corresponds to a specific material, and internal nodes correspond to material family categories. For the process... and process Calculate the shortest path length between the corresponding leaf nodes of the materials used in the hierarchy tree. and the maximum depth of the hierarchy tree ,but:

[0041] ;

[0042] in, This represents the shortest path length of the leaf nodes corresponding to the two materials in the hierarchical tree. The maximum depth of the hierarchical tree, denominator This is an upper bound on the shortest path length between any two leaf nodes in a hierarchical tree. The formula normalizes the hierarchical distance to... The smaller the interval or hierarchical distance, the higher the similarity of the processed material families.

[0043] It should be noted that the above calculation method for the total process similarity score is as follows: for each process of the current product, select the process with the highest process similarity in the process route of the historical product to form the optimal matching pair, sum the process similarity values ​​of all optimal matching pairs and divide by the total number of processes of the current product to obtain the normalized total process similarity score.

[0044] Step 3: Merge the prior cause-effect graphs of the source domain products, perform Bayesian causal structure learning, and generate the posterior cross-process quality causal subgraph of the current product;

[0045] For each source domain product, extract the established cross-process causal relationship directed acyclic graph structure and the effect strength parameters of each causal edge. The causal graph structures of the source domain products are superimposed and fused to generate a fused prior causal graph skeleton and the prior strength distribution of each edge.

[0046] Specifically, for any potential causal edge in the fusion prior causal graph skeleton Potential causal edge Prior confidence According to potential causal edge exist The frequency of each source domain product appearing in the cause-effect graph and the total process similarity score between the corresponding source domain product and the current product are used to determine the following:

[0047] ;

[0048] in, As an indicator function, when causal edges Existing in the In the causal graph of a source domain product, the value is 1 when it is active and 0 otherwise. For the first Total score of process similarity between the source domain product and the current product; For indexing source domain products, Take from 1 to ; The index of the potential causal edge.

[0049] Furthermore, the aforementioned prior confidence level The physical meaning is: in In a single source domain product, under a weighted average with the total score of process similarity as a weight, the causal edge... The frequency estimate of existence. The prior confidence level when a causal edge frequently appears in source domain products with high similarity to the current product's process. Higher prior confidence level indicates a higher confidence level, and vice versa. It is directly used as the prior distribution parameter for the edge existence probability in Bayesian causal structure learning, and combined with the likelihood function of a small amount of sample data of the current product to calculate the posterior existence probability of each edge. .

[0050] Extract workpiece-level full-process feature summary vectors from a small number of workpieces that have completed all processing steps for the current product. Before inputting these workpiece-level full-process feature summary vectors into Bayesian causal structure learning, preprocess the measured values ​​of key quality parameters for each process using Z-score standardization to eliminate the influence of differences in the dimensions and numerical ranges of quality parameters across different processes on the estimation of causal effects. Bayesian causal structure learning is then performed under the constraints of a fused prior causal graph skeleton, using prior confidence levels... As the prior distribution of the edge existence probability, the posterior existence probability of each edge is calculated by combining the likelihood function of a small number of samples. The probability of existence after deletion is below the threshold. Edges and retain those above the threshold Generate a posterior cross-process quality causal subgraph of the current product from the edges.

[0051] Furthermore, threshold The threshold for determining the posterior probability of an edge exists is given, and its value ranges from 1 to 2. Used to control the sparsity of the cause-effect graph structure: threshold A higher threshold value results in fewer causal edges being retained and a sparser causal graph structure; A lower threshold value results in more causal edges being retained and a denser causal graph structure. In practical applications, the threshold... The threshold can be set based on the current product sample size and the prior reliability of the source domain. When the sample size is small, the threshold can be appropriately increased. To suppress spurious causal edges caused by insufficient data.

[0052] For each causal edge retained in the posterior cross-process quality causal subgraph, the effect strength parameter of the corresponding edge in the source domain product is used as the prior mean, and the posterior distribution of the effect strength of each retained edge is updated by Bayesian linear regression. For causal edges... Effect intensity Effect intensity The posterior distribution is:

[0053] ;

[0054] in, This refers to a small number of workpiece data for the current product. For the prior distribution, The values ​​are the prior mean values ​​of the edge effect strength parameters corresponding to the source domain products. For prior variance; It is a likelihood function based on the current product data; This is the index of the causal edge. From this, the posterior mean of the strength of each preserved edge effect is obtained. and posterior variance Posterior variance The uncertainty range corresponding to the strength of the edge effect.

[0055] Furthermore, the specific execution method of the above Bayesian linear regression is as follows: for causal edges... Let causality be defined. The standardized value of the quality parameter of the parent node process is the independent variable, and the causal edge... The standardized values ​​of the quality parameters of the sub-node processes are dependent variables, based on Gaussian priors. Under constraints, using a small number of workpiece data from the current product as observation samples, the effect strength is solved by maximizing the posterior probability. The posterior distribution. Since both the prior distribution and the likelihood function are Gaussian in form, the posterior distribution... Also following a Gaussian distribution, the posterior mean and posterior variance All of them have analytical closed-form solutions, which can be directly calculated without iterative solutions.

[0056] It should be noted that the aforementioned workpiece-level full-process feature summary vector refers to a multi-dimensional vector formed by concatenating the measured values ​​of key quality parameters of each process for each workpiece that has completed all processes, arranged in the order of the processes. In the current early stage of product trial production, when there are only a few workpieces, the dimension of the workpiece-level full-process feature summary vector is equal to the total number of key quality parameters for all processes.

[0057] In this embodiment of the application, when the When overlaying and fusing the causal graphs of multiple source domain products, the consistency of the causal edges in each source domain product's causal graph is also checked. Specifically, if there are causal edges with opposite directions between the same pair of process nodes in different source domain products, the conflicting causal edges are marked as states to be verified. During the Bayesian causal structure learning process, the posterior existence probability of the two directions of the causal edges of the states to be verified is calculated separately, and the direction with the higher posterior probability is selected as the final causal edge direction.

[0058] Step 4: Calculate the conditional Granger causality between equipment anomalies and process quality parameters, and generate a set of cross-layer causal edges;

[0059] For each equipment node in the equipment failure propagation subgraph, all process nodes executed by that equipment node are located in the posterior cross-process quality causality subgraph, and an equipment-process mapping table is established. Before performing the conditional Granger causality test, the time series of equipment anomaly events and the time series of process quality parameter deviations are preprocessed: the equipment anomaly event time series is a multidimensional binary sequence and does not require normalization; the dimensions and numerical ranges of each quality parameter in the process quality parameter deviation time series may differ, so Z-score standardization is used for preprocessing to make each parameter have a comparable numerical scale and eliminate the influence of dimensional differences on the estimation of regression coefficients. For each pair of equipment nodes and process nodes in the equipment-process mapping table, the conditional Granger causality test statistic between the equipment anomaly event time series of that equipment node and the process quality parameter deviation time series of that process node is calculated, where the conditional variable includes the quality parameter time series of all parent node processes of that process node in the posterior cross-process quality causality subgraph to eliminate the confounding effect of inter-process quality transfer effects.

[0060] Based on the conditional Granger causality test statistic and its corresponding significance level, cross-level causal edges that significantly affect process quality parameters due to equipment anomalies are identified. For each cross-level causal edge, the impact delay and effect strength are further calculated. The impact delay is determined by selecting the lag order that maximizes the Granger causality test statistic, and the effect strength is estimated by introducing the equipment anomaly variable into the conditional Granger causality test to estimate the model's explained variance increment. This generates a set of cross-level causal edges, each containing an initial equipment node identifier, a target process node identifier, an impact delay, and an effect strength.

[0061] It should be noted that the aforementioned equipment anomaly event time series refers to expanding the complex cluster of equipment anomalies into a multidimensional binary sequence over discrete time steps according to a time window. Each dimension corresponds to a semantic event type, and is set to 1 if that type of semantic event occurs within that time step, and 0 otherwise. The aforementioned process quality parameter deviation time series refers to the time series formed by arranging the deviations between the measured values ​​and target values ​​of the quality parameters of each workpiece processed in that process according to the workpiece processing time.

[0062] In this embodiment, a regularized vector autoregressive model is used when performing the conditional Granger causality test. This is applied to the time series of equipment anomaly events. Time series of deviations in process quality parameters In condition variables Based on this, the following regularized vector autoregressive model is established:

[0063] ;

[0064] in, This is the current time step; The maximum lag order; For lag order index, Take from 1 to ; These are the autoregressive coefficients of the process quality parameters; This is a vector of influence coefficients on process quality parameters caused by equipment malfunction events. This is the vector of regression coefficients for the condition variables; For the first Time series of device abnormal events at each time step; For the first Deviation of process quality parameters at each time step; For the first Conditional variables at each time step; Indicates vector transpose; This represents the residual term. After Z-score normalization preprocessing... , and All values ​​are dimensionless standardized values, and addition and subtraction can be performed directly on each item. This is achieved by analyzing the influence coefficient vector. Apply Regularization constraints and testing of influence coefficient vectors We can determine whether there is a cross-layer causal relationship by checking if the vector is significantly non-zero.

[0065] Furthermore, the above The purpose of regularization constraints is to mitigate the impact coefficient vector when there are many dimensions of equipment malfunction events but a relatively limited number of effective samples. By imposing sparsity constraints, weak correlation coefficients without real physical meaning are prevented from being overfitted as significant influences, thereby improving the accuracy of cross-layer causal edge identification. The influence coefficient vector is then tested. The way to determine if a vector is significantly non-zero is: in After regularization, the influence coefficient vector is solved. The coefficients corresponding to the non-zero components are subjected to F-test or likelihood ratio test, and the significance level of the test statistic is used to determine whether the cross-level causal relationship is valid.

[0066] Step 5: Assemble the two-layer causal graph structure to generate a two-layer causal graph with complete observational evidence;

[0067] The equipment failure propagation subgraph is used as the upper layer, the posterior cross-process quality causal subgraph is used as the lower layer, and each cross-layer causal edge in the cross-layer causal edge set is used as an inter-layer connection, which is assembled into a two-layer causal graph structure.

[0068] A unified representation of observed variables is established for each node in the two-layer causal graph. The observed variables for the device-layer nodes are multi-dimensional binary vectors corresponding to each semantic event type in the composite abnormal event cluster. If the device-layer node... Each node involves If there are semantic event types, then the device layer's first... Observed variables of each node ,in For the device layer The number of semantic event types involved in each node. The observed variable of the process layer node is the quality parameter deviation vector. If the process layer node... Each node involves The quality parameter, then the process layer Observed variables of each node ,in For the process layer The number of quality parameters involved in each node; before inputting the process-level observed variables into the extended structured Bayesian network model, the process-level... Observed variables of each node The deviation components of each quality parameter are preprocessed using Z-score standardization to eliminate the influence of differences in the dimensions of different quality parameters on the estimation of model parameters. This generates a two-layer causal map with complete observational evidence.

[0069] Step 6: Build an extended structured Bayesian network model on the two-layer causal graph;

[0070] An extended structured Bayesian network model is built upon the two-layer causal graph. The extended structured Bayesian network model includes three types of latent variables and their corresponding conditional probability distributions.

[0071] For each node in the device layer Introducing latent variables of fault sources and propagation of hidden variables Hidden variables of the fault source Indicates whether the device node is an independent fault source; fault source hidden variable. A value of 1 indicates that the device node has a local independent fault. (Propagation and reception of hidden variables) ,in Indicates from the device layer The strength of the propagation effect received by each upstream parent node. Represents device layer node The set of all upstream parent nodes, Represents device layer node The number of upstream parent nodes. Device layer nodes. Observed variables The conditional probability distribution is jointly determined by the latent variables of the fault source and the latent variables of the propagation and reception:

[0072] ;

[0073] in, For parameterized conditional probability functions, For device layer nodes The conditional probability parameter.

[0074] For each node in the process layer Introducing latent variables of process deviation and the transfer and reception of implicit variables between processes Latent variables related to process deviations This represents the deviation component of the quality parameters at this process node caused by the processing conditions of this process. It also represents the implicit variables passed between processes. ,in Indicates from the process level The quality deviation component transmitted from the upstream parent node Represents process layer nodes The set of all parent nodes in the posterior cross-process quality causal subgraph Represents process layer nodes The number of parent nodes in the posterior cross-process quality causal subgraph.

[0075] For cross-layer causal edges, for each connected device layer node and process layer nodes Cross-layer causal edges introduce cross-layer transitive latent variables Passing latent variables across layers Represents device layer node Fault status for process level nodes The influence components of the quality parameters.

[0076] Process layer nodes Observed variables The conditional probability distribution is determined by the three types of latent variables mentioned above:

[0077] ;

[0078] in, For parameterized conditional probability functions, For process layer nodes The conditional probability parameters. It should be noted that the conditional probability parameters of each causal edge in the process layer use the posterior distribution of the effect strength output from the Bayesian causal structure learning in step 3 as the prior parameters, i.e., the process layer nodes. conditional probability parameters The a posteriori mean of the parameter components related to inter-process transfer and posterior variance These are prior constraints.

[0079] It should be noted that the conditional probability function of the above-mentioned device layer nodes... The specific form is as follows: the effects of the fault source latent variable and each propagation and reception latent variable are linearly superimposed and then mapped to the independent occurrence probability of each semantic event type through a logical function. Specifically, for device layer nodes... The The first semantic event type The probability of occurrence of each semantic event type is:

[0080] ;

[0081] ;

[0082] in, It is a logical function, that is , Represents an exponential function, mapping the result of a linear combination to... An interval represents the probability of an event occurring. For the local fault source to the first Influence coefficient of semantic event type; For the first The propagation effect of the upstream device on the first Influence coefficient of semantic event type; This is a bias term. The above influence coefficients... , and bias terms Together they constitute the device layer node conditional probability parameters .

[0083] It should be noted that the conditional probability function of the above-mentioned process layer nodes... The specific form is as follows: It is assumed that the observed variables follow a Gaussian distribution with the mean of a linear combination of the latent variables. Specifically, the process layer nodes... The conditional distribution of the quality parameter deviation vector is:

[0084] ;

[0085] in, The effect intensity coefficient is the effect transferred between processes. The prior distribution is determined by step 3; This indicates that the nodes are connected to the process layer nodes via cross-layer causal edges. The set of all device layer nodes; To observe the noise covariance matrix. Due to the process layer nodes. Observed variables Z-score normalization preprocessing has been performed; latent variables of process-specific biases are excluded. Inter-process transmission and reception of implicit variables and cross-layer transitive of latent variables All are defined in the standardized space, with consistent dimensions, and can be directly added. The above effect intensity coefficients... and observation noise covariance matrix Together they constitute the process layer nodes conditional probability parameters .

[0086] Furthermore, in the extended structured Bayesian network model, device layer nodes... conditional probability parameters and process layer nodes conditional probability parameters All parameters are trained using maximum a posteriori estimation, with the sum of the log-likelihood of the observed variables and the prior log-probability of the parameters as the optimization objective. The Adam optimization algorithm is used for parameter updates. The training input for each device-level node is a multidimensional binary observation vector of that device-level node. The training objective is to maximize the conditional log-likelihood of the observed variables given the latent variables, i.e., to maximize...

[0087] ;

[0088] The sum of the prior log probabilities of the parameters, where For indexes of semantic event types, Take from 1 to The training input for each process-level node is the quality parameter deviation observation vector of each process-level node after Z-score normalization preprocessing. The training objective is to maximize the log-likelihood under the Gaussian conditional distribution, i.e., to maximize...

[0089] The sum of the prior log probabilities of the parameters, where This represents the transpose of a vector. The mean is a Gaussian distribution, and the parameter components related to inter-process transfer are influenced by the posterior mean. and posterior variance Constrained Gaussian prior regularization.

[0090] Step 7: Perform structured variational inference and output multi-source, multi-level joint attribution results;

[0091] Joint posterior inference is performed on an extended structured Bayesian network model using a structured variational inference algorithm. A family of variational distributions is defined. For all latent variables The joint posterior distribution is approximated by the latent variable, where the latent variable It includes the hidden variables of fault sources and propagation and reception of all nodes in the equipment layer, the hidden variables of process deviations and inter-process transmission and reception of all nodes in the process layer, and all cross-layer transmission hidden variables.

[0092] The optimization objective of variational inference is to maximize the lower bound of evidence:

[0093] ;

[0094] in, It is the set of observational evidence for all nodes in the two-layer causal graph; Given all latent variables, the joint likelihood of the observed variables; For the prior distribution of latent variables; Let KL divergence be denoted as KL divergence.

[0095] Furthermore, the lower bound of the aforementioned evidence In the middle, the first item To reconstruct the expected term, measure its value in the variational posterior distribution. The degree to which the observed data is explained by the model; a larger reconstruction expectation term indicates a better fit of the model to the current observational evidence; the second term... The regularization term measures the difference between the variational posterior distribution and the prior distribution of the latent variables. The regularization term serves to constrain the variational posterior from deviating excessively from the prior. The sum of the two terms constitutes the logarithmic evidence. The lower bound of evidence, maximizing the lower bound of evidence. This is equivalent to achieving a balance between fitting the observed data and preserving the prior constraints, thereby obtaining the optimal approximation of the joint posterior distribution of all latent variables.

[0096] By iteratively optimizing the parameters of the variational distribution until the lower bound of evidence converges, an approximate posterior distribution of all latent variables is obtained. From this approximate posterior distribution, the following three types of attribution results are extracted:

[0097] First, the posterior probability of the fault source at each node in the device layer. and the posterior mean of each propagation and reception latent variable Based on the posterior probability of the fault source, sort them from high to low to generate a probability ranking of independent fault sources for each device.

[0098] Second, the posterior distribution of process-specific deviations at each node of the process layer. The process is sorted from largest to smallest based on the posterior norm of its own deviation vector, and the process's own deviation contribution is ranked.

[0099] Third, the posterior distribution of latent variables transmitted across layers. Based on the posterior norm of the latent variables transmitted across layers, sort them from largest to smallest to generate a ranking of the transmission path contribution of equipment failures across layers to process quality.

[0100] All three types of attribution results were calculated in the standardized space. When generating the final attribution report, the posterior mean of each latent variable was restored to the original dimension of the deviation estimate through the Z-score standardized inverse transformation of the corresponding process quality parameter, so as to compare it with the actual process specifications.

[0101] In the embodiments of this application, variational distribution family A structured mean-field approximation is employed. Specifically, the latent variables are divided into three groups according to the hierarchical structure of the two-layer causal graph for approximation: the device-level latent variable group... Hidden variable groups at the process level and cross-level latent variable groups The variational distribution is decomposed into:

[0102] ;

[0103] Within each group, the dependencies between latent variables are preserved, and the groups are treated as mean-field approximate independence. The variational parameters are updated alternately between the three groups until the lower bound of evidence converges.

[0104] Furthermore, the alternating update method for the above three sets of variational parameters is as follows: in each iteration, the parameters are fixed. and ,right The variational parameters are updated using gradient ascent to maximize the lower bound of evidence; then fixed. and ,right The variational parameters are updated using gradient ascent; finally, the values ​​are fixed. and ,right The variational parameters are updated using gradient ascent. These three steps constitute a complete coordinate ascent iteration, which is repeated until the lower bound of evidence is reached. The change between two adjacent iterations is lower than the preset convergence threshold.

[0105] Step 8: Perform propagation path tracing and generate a multi-source, multi-layer joint attribution report with uncertainty quantification;

[0106] The posterior probability of the fault source exceeds the decision threshold. The root cause is traced and the propagation path is tracked.

[0107] Furthermore, the threshold for determination This is the posterior probability threshold for root cause confirmation, and its value range is... Used to distinguish significant fault sources from background noise: Judgment threshold A higher value indicates that only device nodes with a higher posterior probability are identified as independent fault sources, reducing false alarms; the judgment threshold. A lower value will include more candidate nodes in the tracking range, reducing false negatives. In practical applications, the decision threshold... The settings can be comprehensively adjusted based on the workshop's tolerance for false alarms and false negatives.

[0108] For independent equipment failure sources, the propagation path is traced forward from the failure source node within the equipment layer graph, extending to the affected process nodes via cross-layer causal edges. For root causes of deviations within a process itself, the path is traced forward from the process node to the final quality indicator node within the process layer graph.

[0109] For each tracking path, calculate the quantitative contribution of that path to the final quality deviation. Obtain the posterior estimates of the side effect strength along each tracking path sequentially and multiply them to obtain the cumulative effect strength of the tracking path. Normalize the cumulative effect strength of all tracking paths to obtain the contribution ratio of each tracking path to the final quality deviation.

[0110] Simultaneously, the uncertainty intervals of the strength of each side effect at the process level are propagated along the tracking path to calculate the confidence interval of the cumulative contribution. Specifically, for the process layer... The variance of the posterior distribution of the effect strength of each edge in the tracking path composed of causal edges is: The posterior mean of the intensity of each side effect is ,in To determine the total number of causal edges in the tracking path, the approximate variance of the cumulative effect strength of the tracking path is calculated using the error propagation formula:

[0111] ;

[0112] in, To track the first in the path The effect strength of a causal edge To track the indices of causal edges in the path, Take from 1 to ; Effect intensity The posterior mean, Effect intensity The posterior variance.

[0113] Furthermore, the above error propagation formula is derived based on the assumption that the posterior distributions of the strengths of each side effect are independent, using an approximate expansion of the variance of the product. Specifically, for The variance of the product of three independent random variables can be approximately expressed as the square of the product of the means of each variable multiplied by the sum of their relative variances.

[0114] The approximation has a coefficient of variation for the intensity of each side effect. Higher accuracy is achieved when the value is smaller. This yields a confidence interval for the cumulative contribution of the tracking path, which is then appended to the attribution report.

[0115] The report summarizes the probability ranking of independent fault sources of the above equipment, the contribution ranking of process deviations, the contribution ranking of cross-layer transmission paths, and the contribution ratio and confidence interval of each tracing path to generate a multi-source, multi-layer joint attribution report with uncertainty quantification.

[0116] In this embodiment, based on step 8, the following steps are further included: The multi-source, multi-layer joint attribution report is written into the quality management module of the workshop information management system and routed according to root cause type. Equipment failure root causes are routed to the equipment maintenance management module, generating an independent maintenance work order and marking the posterior probability value of the failure source, the propagation impact range, and the contribution ratio of the failure source to anomalies in each downstream equipment. Process deviation root causes are routed to the process management module, generating process parameter adjustment suggestions and marking the confidence interval of causal contribution. Cross-layer impact path information is simultaneously sent to both the equipment maintenance management module and the process management module to coordinate the priority of handling.

[0117] In this embodiment, as the number of prototype workpieces for the new product gradually increases, the incremental data of the newly added workpieces is incorporated into the Bayesian causal structure learning and effect strength update process in step 3. Specifically, the newly added workpiece data is treated as new observations, and a Bayesian incremental update is performed based on the current posterior distribution. The current posterior distribution is used as the prior distribution for the new round of updates, and multiplied by the likelihood function of the newly added data to obtain the updated posterior distribution, continuously reducing the posterior variance of the effect strength, i.e., the uncertainty interval. When the posterior probability of the causal graph structure changes significantly, the reassembly of the two-layer causal graph and the re-execution of parameter updates and variational inference of the extended structured Bayesian network model are automatically triggered.

[0118] The following is an example of an application of the present invention, such as... Figure 2-8 As shown, the implementation process is as follows:

[0119] A digital precision parts processing workshop undertook a project to manufacture a new type of aero-engine blade disk (hereinafter referred to as "the product"). The workshop includes the rough milling process ( ) trial production task. ), finish milling process ( Grinding process ( ) and final inspection process ( The four key processes are handled by the machining center. Machining Center ,grinder and testing station implement. and Shared hydraulic station , It is equipped with an independent hydraulic unit.

[0120] On March 14, 20XX, the quality management module of the workshop information management system issued an alarm: [The following appears to be a separate, unrelated sentence:] Workpiece number [number] had an out-of-tolerance surface profile during the final inspection process. The attribution calculation service was initiated to perform multi-source, multi-level joint attribution for this quality anomaly. As of the time attribution was triggered, the product... A total of 8 workpieces have completed all processing steps. to The historical product database contains historical products with similar craftsmanship. , , Sufficient processing data has been accumulated and cross-process causal relationship diagrams have been established.

[0121] Corresponding to step 1 in the specific implementation, the attribution calculation service extracts the device anomaly event records from 07:00 to 09:30 on March 14, 20XX from the log database, covering... , , , Four devices. After semantic normalization of the original event records, "low hydraulic system pressure" and "decreased hydraulic pump pressure" were uniformly mapped to semantic event types. Map "spindle vibration exceeding limits" to Map "Feed axis position deviation alarm" to Map the "machining dimension out-of-tolerance warning" to .

[0122] right Clustering of abnormal event sequences based on execution time. The typical event response time threshold is 180 seconds (taken from the median time interval of historical consecutive abnormal events). 07:12:05 With 07:13:47 The time interval is 102 seconds, which is less than the 180-second threshold, and therefore belongs to the same complex cluster of abnormal events. ; 07:52:33 With 07:55:18 The time interval is 165 seconds, which is less than the 180-second threshold, and is therefore classified into the composite abnormal event cluster. . The typical event response time threshold is 200 seconds, 07:14:22 With 07:16:09 The time interval is 107 seconds, which is classified into the composite abnormal event cluster. .

[0123] On the device event propagation graph, taking the current alarm device as an example... Extract the reverse reachable subgraph for the endpoint to obtain the subgraph containing Physical propagation path and The subgraph of the shared infrastructure coupling path attaches the composite anomaly clusters of each device node as observation evidence to the corresponding node, as shown in Tables 1 and 2:

[0124] Table 1 Input data for step 1 (original records of device malfunction events)

[0125]

[0126] Table 2 Output data of step 1 (nodes of the equipment failure propagation subgraph with observational evidence)

[0127]

[0128] Corresponding to step 2 in the specific implementation, the attribution calculation service extracts products from the process database. The four process steps are defined and their process parameters are specified, and historical products are retrieved. , , The process route. With Taking the similarity calculation between (fine milling) and the corresponding processes of each historical product as an example, the weighting coefficient is set as follows: , , , ,satisfy .

[0129] by and Corresponding process Taking similarity calculation as an example: both belong to the precision milling process type, All operations employ a five-axis linkage milling method. The overlap rate of the cutting speed specification range is 0.82, the overlap rate of the feed rate specification range is 0.76, and the average value of the overlap rate of the key parameter ranges is taken. Both materials belong to the same subclass of the titanium alloy family, and their path lengths in the material family hierarchy tree are... Maximum depth of hierarchical tree ,but:

[0130] ;

[0131] Substitute into the process similarity formula:

[0132] ;

[0133] ;

[0134] ;

[0135] After calculating the similarity for all process pairs, for each historical product, the best matching pair is selected from the similarity between each process of the current product and each process of the historical product. The sum of the similarity values ​​of all best matching pairs is then divided by the total number of processes of the current product to obtain the normalized total process similarity score. Let... All three historical products were selected as source domain products, and their total process similarity score was directly used as the weight for the prior confidence calculation in step 3. As shown in Table 3:

[0136] Table 3 Output data of step 2 (total score of historical product process similarity)

[0137]

[0138] Corresponding to step 3 in the specific implementation, the attribution calculation service extracts the cross-process causal relationship directed acyclic graph of the three source domain products and calculates the prior confidence for each potential causal edge.

[0139] With potential causal edge ( Taking the rough milling contour deviation as an example (transferring to the finish milling allowance distribution): Does this edge exist in the causal graph? ), There exists (in) ), It does not exist in ( The total similarity scores of the three source domain products are as follows: , , ,but:

[0140] ;

[0141] ;

[0142] ;

[0143] With potential causal edge ( Taking the transfer of surface roughness from fine milling to grinding removal as an example: all three source domain products contain this edge. ,but:

[0144] ;

[0145] For 8 workpieces that have completed all processing steps ( to Extract the feature summary vector of the entire process and perform Z-score standardization on the quality parameter deviations of each process. Set a threshold. By combining the prior confidence with the likelihood function of a small number of samples, the posterior probability of each potential causal edge is calculated. ( If the prior confidence is low and the sample data does not support it, and the posterior probability is below the threshold, the edge is deleted; the other three edges are retained, and the posterior mean of the effect strength of each retained edge is updated by Bayesian linear regression. and posterior variance A posterior cross-process quality causal subgraph is generated, as shown in Table 4:

[0146] Table 4 shows the output data of step 3 (causal edge prior and posterior confidence).

[0147]

[0148] Corresponding to step 4 in the specific implementation method, establish the equipment-process mapping relationship: , , Perform conditional Granger causality tests on each mapping pair, using... Device malfunction event time series (including and (two-dimensional components) and Taking the inspection of the time series of deviations in process quality parameters as an example, the condition variable is... The quality parameter time series (excluding the confusion caused by inter-process transmission).

[0149] The time series of equipment anomaly events is a multidimensional binary sequence and does not require normalization. The time series of process quality parameter deviations is preprocessed using Z-score standardization. Let the maximum lag order be... ,right Apply After regularization, in lag hour of The component is significantly non-zero, F test The value is 0.024, which is below the significance level of 0.05, confirming the existence of a cross-layer causal edge. The effect delay was two time steps (approximately 18 minutes), and the effect strength was 0.318. The test results are summarized in Table 5 below:

[0150] Table 5 shows the output data of step 4 (set of cross-layer causal edges).

[0151]

[0152] Corresponding to step 7 in the specific implementation, structured variational inference is performed on the assembled two-layer causal graph and extended structured Bayesian network model. The variational distribution is approximately decomposed into device-level latent variable sets using a mean-field approximation. Hidden variable groups at the process level and cross-level latent variable groups Three groups, alternating between the three groups to perform coordinate ascent iteration: the last two groups are updated after fixing. Then fix the two sets of updates. Finally, fix the first two groups for updating. The process continues in a loop until the lower bound of the evidence converges.

[0153] The lower bound of evidence after 47 rounds of iteration The change decreased to Below the preset convergence threshold The iteration converges. Three types of attribution results are extracted from the approximate posterior distribution:

[0154] On the equipment layer, Posterior probability of fault source Significantly higher than of ;

[0155] right Propagation of the posterior mean of the latent variables ,show A significant portion of the anomalies stem from Through shared hydraulic stations The spread of.

[0156] In terms of process level, The posterior norm of the process-specific deviation is 0.634 (standardized space). It is 0.287. It is 0.193. The most significant contribution comes from deviations within the process itself.

[0157] Regarding cross-layer transmission, path( The posterior norm is 0.481. path( The posterior norm is 0.356. All three attribution results were calculated in the standardized space and will be restored to the original dimensionless bias estimate in step 8 using the inverse Z-score transformation.

[0158] Corresponding to step 8 in the specific implementation, a judgment threshold is set. . The posterior probability of the fault source is 0.873, which exceeds the threshold, and it is confirmed as an independent fault source. The posterior probability of 0.412 does not exceed the threshold, therefore it is determined to be affected. The impact spreads to downstream equipment rather than being an independent source of failure.

[0159] right Path execution propagation path tracing, the path contains cross-layer edges (Effect intensity 0.521) and inter-process boundary ( , ), ( , ), ( , There are a total of 4 causal edges, namely .

[0160] Posterior mean of cumulative effect strength:

[0161] ;

[0162] The approximate variance of the cumulative effect intensity, where the cross-layer edge The effect strength is directly estimated by the Granger test, with an approximate apparent variance of 0 used in the calculation.

[0163] ;

[0164] ;

[0165] ;

[0166] The standard deviation of the cumulative effect intensity is approximately The corresponding 95% confidence interval is (Lower bound is truncated to 0).

[0167] right Process deviation path (direct connection) ,unilateral The cumulative effect strength of ) is The confidence interval is:

[0168] ;

[0169] After normalizing the cumulative effect intensity of all tracking paths, the contribution ratio of each path is obtained. The posterior mean of each latent variable is then subjected to inverse Z-score normalization transformation of the corresponding process quality parameter. , The profile deviation estimate (mm) is restored to its original dimensions, and a multi-source, multi-level joint attribution report with uncertainty quantification is generated. The report is then routed to the corresponding management module according to the root cause type, as shown in Table 6.

[0170] Table 6 shows the output data from step 8 (summary of multi-source, multi-level joint attribution report).

[0171]

[0172] The data flow throughout the attribution process exhibits a progressive logic, moving from the original event to the structured graph and then to probabilistic inference.

[0173] Step 1: Extract 7 original equipment anomaly event records from the log database. After semantic normalization and temporal clustering, generate 4 composite anomaly event clusters, which are then attached to the corresponding nodes of the equipment fault propagation subgraph, forming an upper-level structure with observational evidence. Step 2: Calculate the product... Based on the total process similarity scores (0.88, 0.81, 0.74) of the three historical products, the source domain product set is determined, and the total similarity score is directly used as the weight for the prior confidence calculation in step 3. Step 3 involves overlaying and fusing the causal graphs of the three source domain products. The prior confidence of the four potential causal edges is calculated using the total similarity score as weight. The posterior existence probability is then calculated by combining this with the standardized full-process feature summary vectors of the eight workpieces. Edges with a posterior probability lower than 0.50 are deleted. The three edges are retained, and the posterior distribution of the effect strength is updated using Bayesian linear regression. and Step 4 constructs a posterior cross-process quality causal subgraph as the lower-level structure, using the equipment-process mapping relationship as the test object. It utilizes the process-level parent node relationship determined in Step 3 as a conditional variable to eliminate confounding, performing a conditional Granger causality test on the equipment anomaly event time series and the Z-score-normalized process quality parameter deviation time series to identify four cross-layer causal edges and their impact delay and effect strength, providing inter-layer connections. Step 5 assembles the above three parts into a complete two-layer causal graph. Step 6 builds an extended structured Bayesian network model on top of it, where the prior constraints of the process-level conditional probability parameters directly come from the output of Step 3. and Step 7 performs structured variational inference. After 47 iterations and convergence, it outputs the posterior distributions of the three types of latent variables, where... The posterior probability of the fault source is 0.873 and The posterior norm of 0.634 for process-specific deviations is a direct input for step 8, determining the root cause and tracing the path. Step 8 addresses deviations exceeding the threshold. The root cause The path tracing is performed, and the confidence interval is calculated using the posterior mean and posterior variance of the effect intensity output in step 3 through the error propagation formula. Finally, the posterior mean in the standardized space is inversely transformed into the original dimensional bias estimate, and a multi-source, multi-level joint attribution report with uncertainty quantification is generated and routed to the corresponding management module.

[0174] It is understood that data preprocessing methods known to those skilled in the art include data cleaning, data transformation, and data reduction. Data transformation includes type conversion and normalization and standardization. Although the dimensions and types of data were omitted in the description of the preceding embodiments, data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite step in data processing. Therefore, the previously described well-known data preprocessing steps were not described independently.

[0175] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An information management method for a digital processing workshop, characterized in that, Includes the following steps: Obtain abnormal event records of all devices within the current fault time window, cluster the execution time of abnormal events to generate composite abnormal event clusters, extract the reverse reachability subgraph on the device event propagation graph and use the composite abnormal event clusters as node observation evidence to generate a device fault propagation subgraph; Clustering of abnormal events by execution time generates composite abnormal event clusters, including: For an abnormal event sequence sorted by timestamp on the same device, the time interval between adjacent events is calculated sequentially. If the time interval is less than the typical event response time of the device, the event is grouped into the same composite abnormal event cluster as the previous event; otherwise, a new composite abnormal event cluster is opened. Before performing time clustering, semantic normalization is performed on the original abnormal event records to uniformly map fault descriptions with different forms of expression but the same meaning to predefined semantic event type codes. Obtain the process routes of current and historical products, calculate the process similarity matrix based on a weighted combination of process type matching degree, processing method consistency, key parameter range overlap rate and processing material family similarity, and screen source domain products; Extract the cross-process causal relationship graphs of products from each source domain, overlay and fuse them to generate a fused prior causal graph skeleton. Use the prior confidence as the prior distribution of the edge existence probability, combine it with a small amount of sample data of the current product to perform Bayesian causal structure learning, and generate a posterior cross-process quality causal subgraph. Establish equipment-process mapping relationships, calculate conditional Granger causality test statistics for each pair of equipment nodes and process nodes, and generate a set of cross-layer causal edges that include the effects of time delay and effect strength; The equipment fault propagation subgraph, the posterior cross-process quality causal subgraph, and the cross-layer causal edge set are assembled into a two-layer causal graph. An extended structured Bayesian network model containing fault source latent variables, propagation and reception latent variables, process self-deviation latent variables, inter-process transmission and reception latent variables, and cross-layer transmission latent variables is established on the two-layer causal graph. Perform structured variational inference to obtain the approximate posterior distribution of all latent variables, and extract the probability ranking of independent fault sources of equipment, the ranking of the contribution of process deviations, and the ranking of the contribution of cross-layer transmission paths. For root causes whose posterior probability of the fault source exceeds the judgment threshold, a propagation path tracing is performed. The posterior estimates of the strength of each side effect are multiplied along the path, and the confidence interval of the cumulative contribution is calculated to generate a multi-source, multi-level joint attribution report with uncertainty quantification.

2. The information management method for a digital processing workshop according to claim 1, characterized in that, The weighted combination calculation of the process similarity matrix based on process type matching degree, processing method consistency, key parameter range overlap rate, and processing material family similarity includes: For the i-th process of the current product and the j-th process of the historical product, the process similarity is calculated as a weighted sum of the process type matching degree, processing method consistency, key parameter range overlap rate, and processing material family similarity multiplied by their respective weight coefficients, where the sum of each weight coefficient is 1. The process type matching degree is a binary index, with a value of 1 when two processes belong to the same process type and 0 otherwise; the processing method consistency is a binary index, with a value of 1 when two processes use the same processing method and 0 otherwise; the key parameter range overlap rate is obtained by calculating the ratio of the intersection length to the union length of the specification range intervals of each key parameter common to the two processes and then taking the average; the processing material family similarity is obtained by normalizing the ratio of the shortest path length of the leaf nodes corresponding to the two materials in the material family classification hierarchy tree to the maximum depth of the hierarchy tree; Based on the process similarity matrix, for each process of the current product, the process with the highest process similarity in each historical product is selected to form the optimal matching pair. The process similarity values ​​of all optimal matching pairs are summed and divided by the total number of processes of the current product to obtain the normalized process similarity total score. The top M historical products with the highest process similarity total scores are selected as source domain products.

3. The information management method for a digital processing workshop according to claim 1, characterized in that, The step of extracting cross-process causal relationship graphs of products from each source domain and overlaying and fusing them to generate a fused prior causal graph skeleton includes: For any potential causal edge in the fusion prior causal graph skeleton, the occurrence of the potential causal edge in the causal graphs of M source domain products is weighted and averaged with the total process similarity score between each source domain product and the current product to obtain the prior confidence of the potential causal edge. If there are causal edges with opposite directions between the same pair of process nodes in different source domain products, the conflicting causal edges are marked as states to be verified. During the Bayesian causal structure learning process, the posterior existence probability of the two directions of the causal edge of the state to be verified is calculated respectively, and the direction with the higher posterior probability is selected as the final causal edge direction.

4. The information management method for a digital processing workshop according to claim 1, characterized in that, The generation of the posterior cross-process quality causal subgraph also includes: Extract the workpiece-level full-process feature summary vector from the workpiece data of the current product that has completed all processes, and preprocess the key quality parameter measurement values ​​of each process using Z-score standardization. Using the prior confidence level as the prior distribution of the edge existence probability, and combining the likelihood function of a small number of samples, calculate the posterior existence probability of each edge, delete edges with a posterior existence probability lower than the threshold, and retain edges with a posterior existence probability higher than the threshold. For each retained causal edge, the effect intensity parameter of the corresponding edge in the source domain product is used as the prior mean. The posterior distribution of the effect intensity of each retained edge is updated by Bayesian linear regression to obtain the posterior mean and posterior variance of the effect intensity of each retained edge. The Bayesian linear regression uses the standardized quality parameter of the parent node process of the causal edge as the independent variable and the standardized quality parameter of the child node process as the dependent variable. Under Gaussian prior constraints, the posterior distribution of the effect intensity is solved by maximizing the posterior probability.

5. The information management method for a digital processing workshop according to claim 1, characterized in that, The calculation of the conditional Granger causality test statistic for each pair of equipment nodes and process nodes includes: A regularized vector autoregressive model is established for the time series of equipment abnormal events and the time series of process quality parameter deviations. The regularized vector autoregressive model takes the process quality parameter deviation as the dependent variable and the lagged terms of the process quality parameter deviation, the lagged terms of the equipment abnormal event time series, and the lagged terms of the condition variable as independent variables. The condition variable is the time series of quality parameters of all parent node processes of the process node in the posterior cross-process quality causal subgraph. An L1 regularization constraint is applied to the influence coefficient vector of equipment abnormal events on process quality parameters. After the regularization solution is obtained, a significance test is performed on the non-zero components in the influence coefficient vector to determine whether the cross-level causal relationship is valid. The effect delay is determined by selecting the lag order that maximizes the Granger causality test statistic, and the effect strength is estimated by introducing the model's explained variance increment after introducing the equipment anomaly variable.

6. The information management method for a digital processing workshop according to claim 1, characterized in that, In the extended structured Bayesian network model: The conditional probability distribution of the observed variables at each node of the equipment layer is determined by linearly superimposing the effects of the fault source latent variables and the propagation and reception latent variables, and then mapping them to the independent occurrence probabilities of each semantic event type through a logical function. The observed variables of each node in the process layer follow a Gaussian distribution with the mean being the product of the process's own deviation latent variable, the inter-process transmission and reception latent variable and the corresponding effect intensity coefficient, and the cross-layer transmission latent variable. The prior distribution of the effect intensity coefficient is determined by the posterior mean and posterior variance of the effect intensity output by Bayesian causal structure learning. The conditional probability parameters of the equipment layer nodes and the process layer nodes are trained using maximum a posteriori estimation, with the sum of the log-likelihood of the observed variables and the prior log-probability of the parameters as the optimization objective.

7. The information management method for a digital processing workshop according to claim 1, characterized in that, The process of performing structured variational inference to obtain the approximate posterior distribution of all latent variables includes: A variational distribution family is defined to approximate the joint posterior distribution of all latent variables. The variational distribution family adopts a structured mean-field approximation, and the latent variables are divided into equipment-level latent variable groups, process-level latent variable groups, and cross-level latent variable groups according to the hierarchical structure of the two-level causal graph. The variational distribution is decomposed into the product of the three groups of variational distributions. The dependencies between latent variables within the group are preserved, and the relationships between groups are processed independently. The optimization objective of variational inference is to maximize the lower bound of evidence, which includes a reconstruction expectation term that measures the degree to which the observed data is explained by the model and a KL divergence regularization term that measures the difference between the variational posterior distribution and the prior distribution of the latent variables. In each iteration, the other two sets of variational parameters are fixed in sequence, and gradient ascent is performed on the variational parameters of the current set. The three sets of alternating updates constitute a complete coordinate ascent iteration. This process is repeated until the change in the lower bound of evidence between two adjacent iterations is lower than the preset convergence threshold.

8. The information management method for a digital processing workshop according to claim 1, characterized in that, The step of multiplying the posterior estimates of the strength of each side effect along the path and calculating the confidence interval of the cumulative contribution includes: For each tracking path, the posterior mean of the edge effect intensity is obtained sequentially along the path and multiplied to obtain the cumulative effect intensity of the tracking path. The cumulative effect intensity of all tracking paths is normalized to obtain the contribution ratio of each tracking path to the final quality deviation. Based on the assumption that the posterior distributions of the strengths of each side effect are independent, the approximate variance of the cumulative effect strength of the tracking path is equal to the square of the product of the posterior means of each side multiplied by the sum of the relative variances of the strengths of each side effect, thereby obtaining the confidence interval of the cumulative contribution of the tracking path. As the number of prototype workpieces for new products increases, the data of the newly added workpieces is used as new observations. Bayesian incremental updates are performed based on the current posterior distribution. The current posterior distribution is used as the prior distribution for the new round of updates and multiplied with the likelihood function of the newly added data to obtain the updated posterior distribution. When the posterior existence probability of the causal graph structure changes significantly, the reassembly of the two-layer causal graph and the parameter updates and re-execution of variational inference of the extended structured Bayesian network model are triggered.

9. An information management system for a digital processing workshop, used to execute the information management method for a digital processing workshop as described in any one of claims 1 to 8, characterized in that, include: The equipment fault propagation subgraph generation module is used to obtain the abnormal event records of all equipment within the current fault time window, cluster the execution time of the abnormal events to generate composite abnormal event clusters, extract the reverse reachable subgraph on the equipment event propagation graph and use the composite abnormal event clusters as node observation evidence to generate the equipment fault propagation subgraph. Clustering of abnormal events by execution time generates composite abnormal event clusters, including: For an abnormal event sequence sorted by timestamp on the same device, the time interval between adjacent events is calculated sequentially. If the time interval is less than the typical event response time of the device, the event is grouped into the same composite abnormal event cluster as the previous event; otherwise, a new composite abnormal event cluster is opened. Before performing time clustering, semantic normalization is performed on the original abnormal event records to uniformly map fault descriptions with different forms of expression but the same meaning to predefined semantic event type codes. The source domain product screening module is used to obtain the process routes of current products and historical products. It calculates the process similarity matrix based on a weighted combination of process type matching degree, processing method consistency, key parameter range overlap rate and processing material family similarity to screen source domain products. The posterior causal subgraph generation module is used to extract the cross-process causal relationship graphs of products from each source domain, overlay and fuse them to generate a fused prior causal graph skeleton. The prior confidence is used as the prior distribution of the edge existence probability. Combined with a small amount of sample data of the current product, Bayesian causal structure learning is performed to generate a posterior cross-process quality causal subgraph. The cross-layer causal edge generation module is used to establish the equipment-process mapping relationship, calculate the conditional Granger causality test statistic for each pair of equipment nodes and process nodes, and generate a set of cross-layer causal edges that include the influence of time delay and effect strength. The two-layer causal graph assembly and modeling module is used to assemble the equipment fault propagation subgraph, the posterior cross-process quality causal subgraph, and the cross-layer causal edge set into a two-layer causal graph, and to establish an extended structured Bayesian network model on the two-layer causal graph that includes fault source latent variables, propagation and reception latent variables, process self-deviation latent variables, inter-process transmission and reception latent variables, and cross-layer transmission latent variables. The variational inference and attribution extraction module is used to perform structured variational inference to obtain the approximate posterior distribution of all latent variables, and extract the probability ranking of independent fault sources of equipment, the ranking of the contribution of process deviations, and the ranking of the contribution of cross-layer transmission paths. The attribution report generation module is used to perform propagation path tracing for root causes whose posterior probability of the fault source exceeds the judgment threshold, multiply the posterior estimates of the strength of each side effect along the path, calculate the confidence interval of the cumulative contribution, and generate a multi-source, multi-layer joint attribution report with uncertainty quantification.

Images