A method for intelligently identifying the root causes of abnormal production yield in probe devices

Abstract

This invention discloses an intelligent method for mining the root causes of abnormal production yield in probe devices, belonging to the field of industrial intelligent analysis technology. The method includes the following steps: acquiring probe final inspection data and historical yield anomaly case data; preprocessing the probe final inspection data and comparing it with a preset threshold to obtain final inspection failure modes; performing correlation mining on historical yield anomaly case data based on the final inspection failure modes to obtain a set of associated sensitive processes; acquiring process production time-series data and constructing process time-series dependency chains; filtering nodes and removing branches in the process time-series dependency chains to obtain process transmission compression paths; using an improved Granger causality test to determine the abnormal transmission of the process transmission compression paths to obtain a set of substantial abnormal nodes; acquiring time-series slice data, inputting it and the set of substantial abnormal nodes into an improved dynamic Bayesian network model, outputting core root cause nodes and node anomaly coupling relationships, and finally generating a conclusion on the root causes of abnormal production yield in probe devices.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent analysis and statistical inference technology, specifically to an intelligent method for mining the root causes of abnormal production yield of probe devices, which comprehensively utilizes association rule mining, time series causal inference and probabilistic graphical modeling techniques. Background Technology

[0002] As a core component for electrical testing of semiconductor chips, the probe device's production process involves multiple continuous steps, including material prefabrication, probe body forming, probe tip finishing, heat treatment, electroplating strengthening, assembly and alignment. The process parameters between these steps are closely related. During mass production, even small parameter deviations in a single step can easily propagate and spread along the process chain, causing various yield abnormalities such as contact impedance fluctuations, coplanarity abnormalities, and probe tip morphology defects.

[0003] Traditional probe-based methods for analyzing the root causes of production yield anomalies involve collecting measured data of process parameters from each production step, directly comparing the measured values ​​with preset standard thresholds, and identifying the steps where parameters exceed the threshold range as the root cause of yield anomalies, thereby locating the root cause of the anomalies.

[0004] However, abnormal signals in the probe production process are often propagated continuously along the process chain and form a multi-process coupling amplification effect. The single-process parameter threshold comparison method cannot identify the abnormal transmission correlation and coupling mechanism between processes. Therefore, it can only locate the abnormal process nodes that exist independently and cannot fully trace the entire path from the generation of the abnormality to its propagation and amplification. Ultimately, this leads to a one-sided result in the determination of the root cause of yield abnormalities, making it difficult to achieve accurate and efficient root cause mining. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent method for identifying the root causes of abnormal production yields in probe devices, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligently identifying the root causes of abnormal production yield in probe devices, comprising the following steps: Step S1: Obtain probe final inspection data and historical yield abnormal case data. Preprocess the probe final inspection data to obtain failure feature data. Compare the failure feature data with a preset threshold to obtain the final inspection failure mode. Based on the final inspection failure mode, perform association rule mining on the historical yield abnormal case data to obtain the associated sensitive process set. Step S2: Obtain the process production time sequence data of the probe device, construct the process time sequence dependency chain based on the associated sensitive process set and process production time sequence data, and filter nodes and remove branches in the process time sequence dependency chain through the reverse depth-first search algorithm to obtain the process transmission compression path; Step S3: Use the improved Granger causality test to determine the abnormal transmission of the process transmission compression path and screen out the set of substantive abnormal nodes; Step S4: Obtain time-series slice data, input the actual abnormal node set and time-series slice data into the improved dynamic Bayesian network model, output the core root cause node and node abnormal coupling relationship, and generate the root cause conclusion of abnormal production yield of probe device based on the core root cause node and node abnormal coupling relationship.

[0007] Preferably, the steps of acquiring probe final inspection data and historical yield anomaly case data, preprocessing the probe final inspection data to obtain failure feature data, and comparing the failure feature data with a preset threshold to obtain the final inspection failure mode include the following specific steps: Acquire probe final inspection data and historical yield anomaly case data; The probe final inspection data is preprocessed to obtain a failure feature dataset; The failure feature dataset is compared with a preset threshold, and the final failure mode is output. Based on the final inspection failure mode, association rule mining is performed on the historical yield anomaly case data to obtain the associated sensitive process set.

[0008] Preferably, the step of performing association rule mining on historical yield anomaly case data based on final inspection failure modes to obtain a set of associated sensitive processes includes: Based on the final inspection failure modes and historical yield anomaly case data, a transaction set is constructed; Generate a candidate set based on the transaction set and calculate the support. Candidate itemsets are filtered based on support to obtain frequent itemsets; Calculate the confidence and lift of each combination in the frequent itemset; Filter combinations that satisfy the confidence threshold and the lift threshold, and extract the processes corresponding to the combinations to form a set of associated sensitive processes.

[0009] Preferably, the acquisition of process production time sequence data of the probe device, and the construction of process time sequence dependency chains based on the associated sensitive process set and process production time sequence data, includes the following steps: Acquire process production timing data of the probe device, wherein the process production timing data includes process execution sequence data and inter-process influence interface data; Basic nodes are extracted from the set of related sensitive processes. Intermediate processes between basic nodes are automatically supplemented as transmission relay nodes according to the fixed process sequence of probe production to form a complete initial node sequence. The continuous adjacent processes in the initial node sequence are used as nodes, the actual execution order of the processes is used as the direction of the directed edges, and the transmission coefficient of the interface between processes is used as the weight of the directed edges to construct a process time dependency chain in the form of a directed acyclic graph. The propagation coefficient of the directed edge weights is calculated using the maximum information coefficient.

[0010] Preferably, the step of filtering nodes and removing branches in the process sequence dependency chain using a reverse depth-first search algorithm to obtain the process propagation compression path is as follows: Calculate the mutual information value between the parameter time-series data and the final inspection failure feature data of each process node in the process time-series dependency chain, and remove process nodes with mutual information values ​​lower than the mutual information threshold to obtain highly correlated process nodes; A reverse depth-first search algorithm is used to traverse and prune the directed acyclic graph composed of highly correlated process nodes. Cut off directed edges and corresponding branches with a conduction coefficient lower than the conduction coefficient screening threshold, remove isolated nodes and branches without continuous conduction association, retain continuous conduction paths where the conduction coefficient between all adjacent nodes is not lower than the conduction coefficient threshold and the mutual information value of each node is not lower than the mutual information threshold, and generate process conduction compression paths.

[0011] Preferably, the step of using the improved Granger causality test to determine abnormal transmission in the process transmission compression path and screening out a set of substantially abnormal nodes includes: Obtain the compression path of the process transmission and the process production timing data of each process; The stationarity test and conditional heteroscedasticity test were performed sequentially on the production time sequence data of each process. For non-stationary time series data, a linear trend term is first fitted. If the absolute value of the slope of the trend term exceeds a preset threshold, the trend term is retained as an independent outlier. If the sequence is still non-stationary after removing significant trend terms, a first-order difference transformation is performed. For time series data exhibiting conditional heteroscedasticity, an improved Granger causality test method based on the autoregressive moving average generalized autoregressive conditional heteroscedasticity model is used to determine anomaly propagation.

[0012] Preferably, the step of using the improved Granger causality test to determine abnormal transmission in the process transmission compression path and screening out a set of substantially abnormal nodes further includes: For each pair of adjacent process nodes in the process transmission compression path, an autoregressive moving average generalized autoregressive conditional heteroscedasticity model containing the parameters of the preceding process is established as the full model, and a restricted model is obtained by removing the relevant terms of the preceding process parameters from the mean equation of the full model. The full model and the restricted model use the same variance equation structure; Solve for the parameters of the full model and the restricted model, and determine whether there is a Granger causal relationship between the anomalies of the parameters in the preceding process and the anomalies of the parameters in the following process based on the likelihood ratio test; Traverse all adjacent process node pairs in the process transmission compression path, retain process nodes with Granger causality, and remove local noise nodes without Granger causality to obtain the set of substantial abnormal nodes.

[0013] Preferably, the step of acquiring time-series slice data, inputting the actual abnormal node set and the time-series slice data into the improved dynamic Bayesian network model, and outputting the core root cause node and the node anomaly coupling relationship, includes the following steps: Obtain the set of substantive anomaly nodes, divide the time sequence into time slices according to the continuous time sequence of the probe production process, and obtain the time slice data corresponding to each substantive anomaly node; Constructing an improved dynamic Bayesian network; The improved dynamic Bayesian network consists of multiple temporal slice Bayesian subnetworks; Nodes in adjacent time slices are connected by directed edges for anomalous state transitions, and subnetworks corresponding to different time slices are allowed to have different state transition probabilities.

[0014] Preferably, the improved dynamic Bayesian network specifically includes: Embedding a coupling amplification transition factor in the calculation of state transition probabilities of dynamic Bayesian networks; The coupling amplification transfer factor is calculated based on the total number of time slices with consecutive anomalies and the standardized anomaly deviation value of the actual anomaly nodes within each time slice. The state transition probability is corrected based on the aforementioned coupling amplification transfer factor; The specific method for correcting the state transition probability is as follows: when the previous time slice is in a normal state, the original state transition probability is used; when the previous time slice is in an abnormal state, the original state transition probability is weighted, corrected, and normalized using the coupling amplification transition factor.

[0015] Preferably, the step of acquiring time-series slice data, inputting the set of substantive anomalous nodes and the time-series slice data into the improved dynamic Bayesian network model, and outputting the core root cause nodes and node anomalous coupling relationships, further includes: An improved dynamic Bayesian network based on modified state transition probabilities performs forward and backward inference to obtain the joint probability of anomalous states in time slices. The abnormal occlusion probability is calculated based on the conditional probability distribution of the improved dynamic Bayesian network, and the abnormal occlusion probability is compared with the occlusion probability threshold to determine whether there is occlusion dependency coupling. The temporal propagation entropy is calculated based on the joint probability of abnormal states in each temporal slice, and nodes with propagation entropy lower than the propagation entropy threshold are removed. Calculate the SHAP contribution value for the remaining nodes after removal, and determine the nodes with an absolute contribution value greater than the contribution threshold as core root cause nodes; The type of coupling relationship is determined based on the temporal sequence and contribution interaction of the core root cause nodes; The final output yields the root cause conclusions of abnormal production yield of the probe device; The root cause conclusions of abnormal production yield of the probe device include the final inspection failure mode, process transmission path, and the type of coupling relationship.

[0016] This invention provides an intelligent method for mining the root causes of abnormal production yield in probe devices. It comprehensively utilizes association rule mining, time series causal inference, and probabilistic graphical modeling techniques, and has the following beneficial effects: (1) By combining the physical threshold determination of final inspection failure characteristics with Apriori association rule mining based on historical cases, and introducing the standardized deviation verification of the measured parameters of the current batch of processes, the method realizes the accurate mapping from the final inspection failure mode to the set of associated sensitive processes. This method fully integrates historical diagnostic experience with the measured representation of the current problem, providing an accurate initial search space for subsequent analysis and effectively avoiding the blindness and computational overhead of full process traversal analysis.

[0017] (2) By supplementing relay nodes to ensure the integrity of the transmission chain topology, the nonlinear transmission relationship between processes is quantified by the maximum information coefficient, and a reverse depth-first search pruning strategy based on the comprehensive weight of nodes is introduced to compress the candidate process range identified in step S1 into a transmission path consisting of 2 to 4 key processes. This path completely preserves the most likely physical transmission chain from the potentially originating abnormal process to the final inspection failure, and eliminates redundant processes and weak transmission branches without direct mechanistic correlation, providing an accurate analysis object for causal testing.

[0018] (3) To address the common non-normal distribution, autocorrelation, and conditional heteroscedasticity characteristics of probe production process parameters, an improved Granger causality test method based on the ARMA-GARCH model is adopted. Through explicit modeling of fluctuation clustering, dynamic BIC automatic order selection, fuzzy interval-assisted judgment, and transmission chain reconstruction, the abnormal transmission causal relationship between processes is accurately determined. This method effectively eliminates local noise nodes that only have parameter fluctuations but no cross-process transmission ability, and selects substantial abnormal nodes with real causal transmission ability, providing a clean and accurate node foundation for subsequent coupling analysis.

[0019] (4) By relaxing the time homogeneity assumption of traditional dynamic Bayesian networks, it allows different time slices to correspond to independent state transition probability matrices and embeds coupling amplification transition factors based on historical case regression analysis. This accurately quantifies the exponential amplification effect of small deviations in probe production along the process chain. Combined with forward-backward bidirectional reasoning, transfer entropy cross-validation and anomaly masking judgment, it fully restores the entire path of anomaly propagation from the source process along the production time sequence, coupling amplification to final inspection manifestation, and adapts to the non-steady-state scenarios of parallel production of multiple probe models and rapid line changeover.

[0020] (5) By decomposing the contribution of SHAP and improving the quantitative judgment rules of dynamic Bayesian networks, the abnormal coupling relationships between three types of processes, namely sequential dependency, amplification dependency and occlusion dependency, are accurately identified. Combined with the priority handling strategy for multiple condition conflicts and the fallback mechanism for extreme scenarios, the accurate location of the core root cause node and the reliable judgment of the coupling type are realized. Finally, the hierarchical root cause conclusions covering the final inspection explicit failure layer, the process transmission layer and the root cause triggering layer are output, providing a clear decision basis for the precise process control on the production site.

[0021] (6) This invention integrates association rule mining, information theory measurement, time series causal inference and probabilistic graphical modeling into a four-stage progressive analysis architecture of process delineation, transmission path compression, causal verification and screening and coupling modeling and positioning. This achieves automated, accurate and interpretable mining of the root causes of abnormal production yield of probe device, and significantly improves the completeness and reliability of abnormal root cause determination. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the steps of an intelligent method for mining the root causes of abnormal production yield of a probe device proposed in this invention. Figure 2 This is a step hierarchy diagram for obtaining the associated sensitive process set in the intelligent mining method for the root cause of abnormal production yield of probe device proposed in this invention; Figure 3 This is a hierarchical diagram of the steps in the intelligent mining method for identifying the root causes of abnormal production yield of probe devices proposed in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-3 This invention provides a technical solution: a method for intelligently identifying the root causes of abnormal production yield of probe devices.

[0026] Step S1: Obtain probe final inspection data and historical yield abnormal case data. Preprocess the probe final inspection data to obtain failure feature data. Compare the failure feature data with a preset threshold to obtain the final inspection failure mode. Based on the final inspection failure mode, perform association rule mining on the historical yield abnormal case data to obtain the associated sensitive process set. The final inspection data and historical yield anomaly case data of probes are acquired. The final inspection data of probes are collected through high-precision electrical performance testing equipment and three-dimensional morphology testing equipment at the final inspection station of the probe production process. The acquisition frequency is one set of data every 10 seconds. The data storage format is floating-point numerical value. The data covers six types of failure features: contact impedance fluctuation data, coplanarity anomaly data, tip morphology inconsistency data, elastic recovery deviation data, local coating failure data, and indentation anomaly concentration data. Temporal features, statistical features, and distribution features are extracted for each of the six types of failure features to form a high-dimensional characterization. Among them, contact impedance fluctuation data corresponds to 48-dimensional features, coplanarity anomaly data corresponds to 42-dimensional features, tip morphology inconsistency data corresponds to 50-dimensional features, elastic recovery deviation data corresponds to 38-dimensional features, local coating failure data corresponds to 40-dimensional features, and indentation anomaly concentration data corresponds to 38-dimensional features. After the six feature dimensions are accumulated, the feature dimensions of a single batch of data are 256 dimensions.

[0027] The collected final detection data is preprocessed. First, 3D processing is used. The principle is to identify and remove noisy data caused by equipment acquisition failures. Specifically, for any feature, collection points whose values ​​exceed the mean ± 3 standard deviations of the normal production data for the past 3 months are removed. Secondly, values ​​exceeding the normal range (e.g., contact impedance greater than 50 milliohms, coplanarity greater than 5 micrometers) but conforming to process failure patterns are identified as extreme failure samples, retained, and included in the failure feature dataset. This invention uses the Z-score standardization method to preprocess all feature data; the calculation formula is as follows: ,in, This is the preprocessed failure characteristic data. This is the average of this feature across the past three months of normal production data. To correspond to the standard deviation, the standardized data have a mean of 0 and a standard deviation of 1.

[0028] Based on the physical thresholds set in the probe production and testing standards, the pre-processed six types of failure characteristic data are directly compared with their respective standard thresholds. The failure type corresponding to the item exceeding the standard is taken as the final failure mode of this analysis.

[0029] If two or more final inspection failure modes are identified in the current batch, each failure mode is used as an index to independently perform subsequent association rule mining to obtain the associated sensitive process subsets corresponding to each failure mode. All subsets are merged, and duplicate processes are removed to form a unified initial associated process set. During merging, if the same process appears in multiple subsets, the maximum confidence of the association rules in different subsets is taken as the final confidence of the process in the current batch.

[0030] Historical yield anomaly case data is acquired. This case data is exported from the Manufacturing Execution System and includes historical final inspection data, failure type labeling information, and corresponding problematic process records. The historical case database is automatically updated daily at 24:00. During the update, duplicate cases are automatically removed based on the production batch number and final inspection data fingerprint. New cases must be reviewed by process engineers using samples, with a sampling ratio of no less than 10% and a review accuracy rate of no less than 95%. Cases with incorrect labeling are removed from the database. The total size of the case database is maintained between 5,000 and 10,000 records. When the total exceeds 10,000 records, the earliest 1,000 cases are automatically deleted in chronological order of entry into the database.

[0031] Considering that a single historical case may be associated with multiple failure modes and multiple problematic procedures simultaneously, the data needs to be standardized and decomposed to adapt to the input requirements of association rule mining algorithms. Specifically, if a historical case contains M failure modes and N problematic procedures, it is decomposed into M... There are N independent transaction records, each consisting of a failure mode and an independent problem procedure. The transaction set required by the Apriori algorithm is constructed based on this rule.

[0032] Traverse the transaction set and calculate the support of all candidate 1-itemsets (i.e., combinations of {failure modes, single processes}). Considering that in the diagnosis of abnormal yield of probes, the strong correlation between specific failure modes and specific processes often has the distribution characteristics of low frequency and high confidence, the confidence threshold is set to 0.85 to ensure the reliability of the rules. To avoid missing low-to-medium frequency rules with diagnostic value, the support threshold is set to 0.05 (i.e., appearing no less than 250 times in 5000 cases) to achieve a balance between rule coverage and diagnostic value. The lift threshold is set to 1.2 to ensure that the positive correlation of the rules is significantly better than the random level. Candidate combinations with support below 0.05 are eliminated to obtain frequent 1-itemsets. Based on the frequent 1-itemsets, frequent 2-itemsets and 3-itemsets are iteratively generated and screened.

[0033] For each frequently obtained itemset, generate a failure mode in the form of "failure mode". "Single process" or "failure mode" For each generated rule in the "process combination" association rules, its support, confidence, and lift are calculated. The support calculation formula is as follows: ; in, This indicates the specific final inspection failure mode identified by the aforementioned steps. Represents a set of one or more processes. Indicates when a failure mode occurs At that time, it is associated with the process set. The rules, This indicates the support level of the association rule. This indicates the number of historical cases where the corresponding failure mode and process set occur simultaneously. This represents the total number of cases in the historical yield anomaly case library.

[0034] The confidence level is calculated using the following formula: ; in, Indicates when a failure mode occurs At that time, it is associated with the process set. The confidence level of the rule, This indicates the total number of cases in the historical data where the corresponding failure mode occurred.

[0035] Calculate the ratio of confidence level to the probability of a randomly occurring sensitive process to obtain the lift. Select strong association rules with a confidence level of not less than 0.85 and a lift of not less than 1.2, extract the processes corresponding to these rules, and form an initial set of associated processes.

[0036] If no valid association rules that meet the above threshold conditions are obtained after screening, the association rule screening step is skipped, and all six processes of probe production—material prefabrication, needle body forming, needle tip finishing, heat treatment, electroplating strengthening, and assembly leveling—are directly input into step S2 as a set of associated sensitive processes to start the root cause tracing process across the entire process.

[0037] Furthermore, in order to combine historical statistical rules with the actual production status of the current batch, obtain the measured process parameters of each process in the initial associated process set for the current batch, calculate the standardized deviation value between the parameters and the corresponding standard process parameters, set the deviation threshold to 0.10, remove processes with absolute deviation values ​​lower than this threshold, and take the remaining processes with significant deviations as the associated sensitive process set directly related to the current failure.

[0038] Finally, this step outputs a structured set of associated sensitive processes, which serves as the search space for step S2. Each record in this set contains the following information: process identifier, associated final inspection failure mode, and confidence level of the association rule.

[0039] Through the above process, this step comprehensively utilizes final inspection failure mode determination based on physical thresholds and association rule mining based on historical cases, combined with deviation verification of measured parameters of the current batch's processes. Ultimately, it obtains a set of sensitive processes highly correlated with the specific failure modes of the current batch and verified by measured data. This set integrates historical diagnostic experience with the measured characterization of the current problem, providing a precise and reliable initial analysis scope for the subsequent step S2 to construct the anomaly propagation path. It effectively avoids the blindness and computational overhead of examining all processes, serving as a reliable foundational input for subsequent path compression and causal determination.

[0040] Step S2: Obtain the process production time sequence data of the probe device, construct the process time sequence dependency chain based on the associated sensitive process set and process production time sequence data, and filter nodes and remove branches in the process time sequence dependency chain through the reverse depth-first search algorithm to obtain the process transmission compression path; Obtain the structured, associated sensitive process set output from step S1. Each record in this set includes a process identifier, the associated final inspection failure mode, the confidence level of the association rule, and the standardized deviation value of the current batch. Simultaneously, obtain the process production sequence data for the probe device, including the sequence data of process execution and the interface data of inter-process influence. The sequence data records the execution time of six production processes: material prefabrication, needle body forming, needle tip finishing, heat treatment, electroplating strengthening, and assembly leveling. Each process record includes the process number, execution timestamp, and measured values ​​of process parameters. The inter-process influence interface data system defines all physical relationships between adjacent processes, specifically including: the uniformity of hardness of the pre-made blank corresponding to the forming pressure stability parameter of the needle body forming; the outer diameter deviation parameter of the needle body forming corresponding to the cutting allowance parameter of the needle tip finishing; the consistency parameter of the tool tip radius of the needle tip finishing corresponding to the heating rate adaptability parameter of the heat treatment process; the uniformity of hardness of the heat treatment corresponding to the coating adhesion parameter of the electroplating strengthening process; the uniformity of the coating thickness of the electroplating strengthening process corresponding to the reference offset correction parameter of the assembly leveling; and the reference offset parameter of the assembly leveling corresponding to the coplanarity parameter of the final inspection. The selection rule for the core transmission parameters is as follows: from the above correspondences, select the key quality characteristic parameters of the output end of the previous process and the input process control parameters of the subsequent process that are directly affected by them to form parameter pairs, which are used to calculate the transmission coefficient.

[0041] Using all the associated sensitive processes output in step S1 as base nodes, and following the fixed process sequence of probe production (material prefabrication, needle body forming, needle tip finishing, heat treatment, electroplating reinforcement, assembly, and leveling), intermediate processes between the base nodes are automatically added as relay nodes to form a complete initial node sequence. When constructing the initial node sequence, each process node is assigned a comprehensive weight. The weight calculation formula is: Node Comprehensive Weight = Association Rule Confidence. Standardized deviation value for the current batch.

[0042] If there are defined core transmission parameters between adjacent process nodes, a directed edge connection is established with the actual execution order of the processes as the direction of the directed edge; if there are no defined core transmission parameters between adjacent nodes, it is determined that the edge has no direct physical transmission relationship and no directed edge connection is established. If multiple independent sub-chains are formed as a result, all sub-chains are retained and used as independent candidate transmission paths.

[0043] For each directed edge, its transmission coefficient is calculated. The transmission coefficient is quantified using the maximum information coefficient (MIC), which can simultaneously measure linear and nonlinear correlations. The value ranges from 0 to 1, with a larger value indicating a stronger influence of the preceding process on the following process. The formula for calculating the MIC is: ; in, Let X and Y be the transmission coefficients. and Divide variables X and Y into equal parts a and b, respectively. The mutual information value after partitioning is n, which is the number of samples. The sample size of the single process time series data is consistent with the output of the current abnormal batch. If the sample size of the current batch is less than 100 groups, the corresponding process data of the normal production batch of the same model in the past 3 months will be supplemented to make the total sample size not less than 100 groups.

[0044] After assigning directed edge weights, the mutual information value between the time-series parameter data and the final inspection failure characteristic data of each process node is calculated. For each process node, the time-series data of its core transmission parameters and major process parameters are selected, and the mutual information value is calculated one by one with the final inspection failure characteristic data. The maximum value of all mutual information values ​​is taken as the measure of the nonlinear correlation strength between the process node and the final inspection failure. The mutual information calculation formula is: ; in, This represents the mutual information value between process parameters and final inspection failure. These are the key parameters and time-series variables for process nodes. For final inspection failure characteristic variables, This represents the joint probability density of process parameters and final inspection failure. The marginal probability density of the process parameters. The marginal probability density represents the final inspection failure characteristics.

[0045] The preset mutual information filtering threshold is 0.15, which is determined based on the minimum effective correlation value of historical anomaly propagation. Process nodes with mutual information values ​​lower than the mutual information threshold are removed to obtain highly correlated process nodes.

[0046] A reverse depth-first search algorithm is used to traverse and prune the directed acyclic graph of each candidate transmission path. The algorithm starts from the last process node of the process flow (i.e., the assembly and leveling process node) and performs a depth-first search along the reverse direction of the directed edges (i.e., upstream of the process flow). The maximum search depth is set to 5. During the traversal, a transmission coefficient threshold of 0.2 is set. When traversing to any node, if the transmission coefficient of the directed edge connected to the node is lower than the threshold, the directed edge is directly cut and the corresponding branch is removed; if the node is an isolated node with no continuous transmission connection, the isolated node is directly removed; if there is any combination of nodes without continuous transmission in the branch path, the branch path is directly removed, and only continuous transmission paths where the transmission coefficient between all adjacent nodes is not lower than the transmission coefficient threshold and the mutual information value of each node is not lower than the mutual information threshold are retained.

[0047] During the pruning process, if the difference in mutual information value or the difference in transmission coefficient between two nodes is less than 0.05, the node with the higher overall weight is retained first. If the number of remaining nodes after pruning is less than 2, a fallback strategy is adopted: the transmission coefficient threshold is lowered by 0.05 and the mutual information threshold is lowered by 0.03 for re-selection; if there are still less than 2 nodes after re-selection, the newly entered nodes after the threshold reduction and the two nodes with the highest overall weight among the original highly correlated process nodes are retained to form a transmission path; if the total number of highly correlated process nodes before the reduction is less than 2, all highly correlated process nodes are directly used as effective transmission paths.

[0048] If multiple independent transmission paths are obtained after pruning (with no shared nodes between the paths), all paths are retained, numbered separately, and used as an independent set of process transmission compression paths.

[0049] The final result is a structured process transmission compression path. Each path is a sequence of nodes arranged according to the execution order of the processes. Each node in the path contains the following information: process identifier, node comprehensive weight, mutual information value with final inspection failure, transmission coefficient with the preceding node, and standardized deviation value of the current batch. If there are multiple independent paths, they are distinguished by path numbers. This structured path data is directly used as the input for step S3.

[0050] Through the above process, this step deeply integrates the statistically correlated process set derived from historical cases with the actual process sequence flow and physical impact interface relationships in the factory. By combining reverse depth-first search with multi-dimensional pruning, it achieves precise mechanistic-level compression of the anomaly propagation chain. The final generated process propagation compression path eliminates redundant processes and weak propagation branches that have no direct mechanistic connection with the final inspection failure, while fully preserving the most likely physical propagation chain from the potentially originating anomaly process to the final inspection failure. This path serves as the precise analysis scope for subsequent improvements to the Granger causality test, laying the foundation for accurately determining the causal relationship of anomaly propagation between processes.

[0051] Step S3: Use the improved Granger causality test to determine the abnormal transmission of the process transmission compression path and screen out the set of substantive abnormal nodes; The input data for this step is the set of process transmission compression paths output by step S2. If step S2 outputs multiple independent paths, then all the following inspection processes are executed for each path, and each path independently outputs its corresponding subset of substantive abnormal nodes. Finally, the data are merged and deduplicated to form a unified set of substantive abnormal nodes. The time series data sample size for a single process is consistent with that of step S2, that is, consistent with the output of the current abnormal batch. If the sample size is less than 100 sets, then the corresponding process data of the normal production batch of the same model in the past 3 months are added to make the total sample size no less than 100 sets.

[0052] First, the production time series data for each process were preprocessed, and then stationarity and conditional heteroscedasticity tests were performed sequentially. Before the tests, the skewness, kurtosis, and Jarque-Bera statistic of each series were calculated to preliminarily determine the distribution characteristics. The stationarity test adopted the Augmented Dickey-Fuller (ADF) test method, with the lag order automatically determined according to the Bayesian Information Criterion (BIC). The maximum lag order was set to 15. At a significance level of 0.05, if the test statistic was less than the critical value, the null hypothesis of non-stationarity was rejected, and the time series data was determined to be a stationary series.

[0053] For time series data that are determined to be non-stationary by the ADF test, their linear trend term is first fitted. If the absolute value of the trend term's slope exceeds 0.05 (indicating significant parameter drift), the trend term is retained as an independent outlier and included in subsequent analysis along with the original series. First-order differencing is only performed if the series remains non-stationary after removing the significant trend term. If the series is still non-stationary after first-order differencing, the node is marked as statistically unstable and not included in subsequent causal tests, but its bias information is retained for reference in step S4.

[0054] The conditional heteroscedasticity test was performed using the autoregressive conditional heteroscedasticity Lagrange multiplier (ARCH-LM) test method. The lag order was automatically selected from 1 to 10 by the BIC criterion, and the significance level was 0.05.

[0055] If the ARCH-LM test accepts the null hypothesis (i.e., there is no conditional heteroscedasticity), then the traditional vector autoregression (VAR) model is used for Granger causality testing; if the null hypothesis is rejected (conditional heteroscedasticity exists), then the improved Granger causality test method based on the ARMA-GARCH model is used.

[0056] Construct an ARMA-GARCH model, which consists of a mean equation and a variance equation. The mean equation is as follows: ; in, This represents the process parameter value for the subsequent process at the current moment. This is the constant term in the mean equation. The coefficients of the autoregressive term for the parameters of the subsequent process are given. For lag The process parameter values ​​for the subsequent process. The coefficient of the lagged term of the exogenous variable from the preceding process. For lag The parameter values ​​of the preceding process at any given time. For the random disturbance term at the current moment, This represents the autoregressive lag order of the parameters for the subsequent process. The lag order of the exogenous variables in the preceding process. The lag order of the moving average. , and By the BIC criteria within their respective preset ranges ( [1, 15], [1, 5], The optimal value is automatically selected within [1, 10]). The coefficient of the moving average term. For lag The random perturbation term at time step.

[0057] The variance equation, also known as the generalized autoregressive conditional heteroscedasticity equation, is calculated using the following formula: ; in, Let V be the conditional variance of the disturbance term at the current moment. For the constant term of the variance equation, The coefficients of the autoregressive conditional heteroscedasticity term are given. For lag The squared value of the disturbance term at time step. Let be the order of the autoregressive conditional heteroscedasticity term. Let the order of the generalized autoregressive conditional heteroscedasticity term be . and The optimal combination is automatically selected within the range {0, 1, 2} using a grid search combined with the BIC criterion, with the default value set to (1, 1). The coefficients of the generalized autoregressive conditional heteroscedasticity term are given. For lag Conditional variance at time.

[0058] The model parameters are solved by maximum likelihood estimation. After estimation, the Ljung-Box white noise test (10 lags) of the standardized residuals is used to verify the fitting effect. If the model fails to converge or the residuals are autocorrelated (P<0.05), the lag array is automatically adjusted and re-estimated until the test is passed.

[0059] Based on this model, two sets of comparative models were constructed. The first set is a full model including parameters from the preceding process, and the second set is a restricted model excluding parameters from the preceding process. The variance equations, i.e., the GARCH components, of the two sets of models have completely identical structures and parameters. Only the lagged terms of exogenous variables were removed from the mean equation of the restricted model. With the remainder unchanged, calculate the maximum likelihood function values ​​for both models. and Construct the asymptotic likelihood ratio statistic: ; in, This is the asymptotic likelihood ratio statistic. The maximum likelihood function value for the constrained model. This represents the maximum likelihood function value for the entire model.

[0060] The asymptotic likelihood ratio statistic follows a chi-square distribution with L degrees of freedom. At a significance level of 0.05, compare the LR with the chi-square critical value with L degrees of freedom. Comparison, when LR is in When dealing with fuzzy intervals, an auxiliary rule is introduced: if the comprehensive weight of the node is ≥0.5 and the mutual information value with the final failure is ≥0.12, it is determined to pass the causality test; otherwise, it is determined to fail. If it exceeds the upper limit of the fuzzy interval, it passes directly; if it is below the lower limit, it fails directly.

[0061] Causality tests are performed in descending order of the overall weight of the nodes. If a node in the middle of the path fails the test, it is identified as a local noise node and removed. At the same time, the causal relationship between the node before and after it is automatically established, and the cross-node transmission effect is evaluated.

[0062] After traversing all adjacent relationships, local noise nodes that fail the test are removed, marked, and retained for subsequent auxiliary judgment. If all adjacent node pairs of a certain path fail the test, all nodes of that path are retained and marked with a causal strength of 0, and included as nodes to be verified. Subsets of all paths are merged and deduplicated. If the final node set contains only one node, the adjacent process node with the closest comprehensive weight is added from the path in step S2 to ensure that there are no fewer than two nodes.

[0063] The final output is a structured set of substantive anomaly nodes. Each node contains: process identifier, node comprehensive weight, mutual information value with final inspection failure, transmission coefficient with preceding node, Granger causality test LR statistic, causality strength (1=pass, 0=to be verified), and whether it is a local noise node label. This dataset is directly used as the input for step S4.

[0064] Through the above process, this step addresses the non-normality, autocorrelation, and conditional heteroscedasticity characteristics commonly found in probe procedures. By explicitly modeling fluctuation clustering within the ARMA-GARCH framework, robust causal test inferences are obtained. Compared to the Granger test under the traditional VAR framework, the improvement of this invention lies in providing a more reliable likelihood ratio statistic under heteroscedasticity, avoiding spurious causal inferences. The final output set of substantially anomalous nodes eliminates local noise without propagation capabilities and retains nodes with statistically significant causal propagation capabilities, providing a clean and reliable node foundation for subsequent coupling analysis of dynamic Bayesian networks.

[0065] Step S4: Obtain time-series slice data, input the actual abnormal node set and time-series slice data into the improved dynamic Bayesian network model, output the core root cause node and node abnormal coupling relationship, and generate the root cause conclusion of abnormal production yield of probe device based on the core root cause node and node abnormal coupling relationship.

[0066] The input data for this step is the structured set of substantive anomaly nodes output from step S3. Each node includes: process identifier, node comprehensive weight, mutual information value with final inspection failure, transmission coefficient with preceding nodes, Granger causality test LR statistic, causality strength, and whether it is a local noise node. Simultaneously, time-series slice data corresponding to each process is obtained. The parameter data of a single slice is 8-dimensional, specifically including process parameter deviation, surface quality indicators, dimensional accuracy data, elastic performance data, coating performance data, assembly positioning data, heat treatment uniformity data, and leveling accuracy data.

[0067] In this invention, the time-series slices are based on the complete execution cycle of the six processes in probe production, forming six consecutive time-series slices. These slices correspond sequentially to material prefabrication and assembly leveling according to the process order. For slices corresponding to non-substantial anomaly nodes, their observation data are retained, but their state variables are only used as background information for network inference and are not included in the core root cause analysis.

[0068] Step S3 ensures that the number of substantially abnormal nodes is no less than two. If, due to extreme circumstances, there are still fewer than two, the coupling analysis in this step is skipped, the unique node is identified as the "single-node instability" root cause, and the conclusion is output directly. Otherwise, the following process is followed.

[0069] An improved dynamic Bayesian network is constructed. Traditional dynamic Bayesian networks assume that all time slices share the same state transition probability matrix (i.e., the time homogeneity assumption), which cannot adapt to the non-steady-state scenarios of parallel production of multiple probe models and rapid line changeover. The improvement of this invention lies in allowing different time slices to correspond to independent state transition probability matrices. The state transition probability matrix A between adjacent time slices is dynamically initialized based on the conduction coefficient and causality strength obtained in S3 for the corresponding process of that slice. The calculation formula is as follows: ; in, For the conduction coefficient, The causal strength is derived from the node structured information output by S3, and the transition probabilities of different slice pairs are independent of each other. There is no mandatory sharing of parameters, thus adapting to non-steady-state production conditions.

[0070] The initial state probabilities of the improved dynamic Bayesian network (i.e., the anomaly probabilities of each node in the first time slice) are initialized using the node causality strength output by S3: ; ; in, For the initial state probability, the initial abnormal probability of the local noise node marked by S3 is uniformly set to 0.1. In the dynamic Bayesian network, it participates in inference as a pure observation node. Its state is directly determined by the comparison between the measured process parameters of the current batch and the standard threshold. If it exceeds the threshold, it is an abnormal state 1; otherwise, it is a normal state 0. It does not participate in the learning and updating of the state transition probability and is only used to assist in verifying whether the shading effect may be caused by the interference of noise nodes.

[0071] A coupling amplification transfer factor is embedded in the state transition probability calculation stage of the network to quantify the amplification effect of the superposition of small anomalies within continuous time slices. The formula for calculating the coupling amplification transfer factor is as follows: ; in, For coupling amplification transfer factor, The basic transmission coefficient is fixed at 1.2. This value was determined based on regression analysis of 5000 historical anomaly cases. It can best fit the cumulative amplification law of abnormal deviations along the process chain in probe production, with an average fitting error of less than 5%. This represents the number of time-series slices that have continuously detected abnormal states up to the current slice, dynamically calculated during the network's forward inference process. The standardized anomaly deviation value is the value of the actual anomaly node in the s-th time slice. The deviation value is obtained by dividing the absolute value of the difference between the measured parameter of the process and the standard parameter by the standard parameter.

[0072] It should be noted that the specific method of embedding the coupling amplification transition factor is as follows: the time slice state is divided into normal state (denoted as 0) and abnormal state (denoted as 1). Coupling amplification correction is only enabled when the current slice state i is an abnormal state; when the current slice state i is a normal state, the original state transition probability is directly used. The formula for the corrected state transition probability is: ; in, This is the probability that the (t+1)th time slice will transition to an abnormal state j when the t-th time slice is in an abnormal state i after correction. The original state transition probability without embedded coupling amplification transition factor. Let be the abnormal state variable for the t-th time slice; Let i be the abnormal state variable of the (t+1)th time slice, i be the abnormal state number of the tth time slice, and j be the abnormal state number of the (t+1)th time slice. After correction, if the probability value is greater than 1, it is set to 1. At the same time, the probability values ​​of all j under the same time slice are normalized. The state transition probability after embedding this factor can realize the low probability propagation of a single time slice abnormality and the exponential amplification propagation of continuous multi-slice abnormalities, accurately matching the failure characteristics of the cumulative amplification of small deviations in probe production.

[0073] An improved dynamic Bayesian network based on an embedded coupling amplification transfer factor performs bidirectional forward and backward inference. Forward inference implements anomaly probability filtering from the initial slice to the last slice. The forward filtering probability calculation formula is as follows: ; in, Let be the forward filtering probability that the t-th time slice is in an abnormal state i. Let be the abnormal state variable for the t-th time slice. The observed parameter data for the t-th time slice.

[0074] Backward inference smooths the anomaly probability from the end slice to the initial slice. The formula for calculating the backward smoothing probability is: ; in, Let be the backward smoothing probability that the t-th time slice is in an abnormal state i. Given the total number of time-series slices, the joint probability of anomalous states in time-series slices is obtained by combining the results of forward filtering and backward smoothing, thus completing the full path tracing of anomalous propagation across time-series slices.

[0075] Combining the forward filtering probability a:(i) and the backward smoothing probability β:(i), we obtain the joint posterior probability that the t-th time slice is in an anomalous state i: ; in, Let be the joint posterior probability that the t-th time slice is in an abnormal state i. This process uses observation data from all time-series slices to complete the full path tracing of anomaly propagation across time-series slices.

[0076] In the probability calculation process of bidirectional inference, the anomaly occlusion probability is calculated synchronously based on the network's conditional probability distribution. This probability is used to determine the effect of subsequent temporal slice anomaly features overshadowing the weak anomaly features of preceding temporal slices in the final inspection observation. This belongs to the observation feature level occlusion rather than the temporal causal level occlusion. The occlusion probability calculation formula is as follows: ; in, This represents the probability of abnormal occlusion. This represents the marginal probability of anomalies occurring in the preceding temporal slice. This represents the conditional probability of an abnormal state in a preceding time-series slice given that an abnormal state exists in a subsequent time-series slice.

[0077] Calculate the probability of abnormal occlusion Afterwards, if If the threshold is >0.3, further verify whether the node corresponding to the preceding slice is a local noise node marked by S3. If it is not a noise node, it is determined that there is occlusion dependency coupling. If it is a noise node, the occlusion probability threshold is relaxed to 0.5 to improve the robustness of the judgment. This mechanism can effectively identify typical occlusion problems such as electroplating abnormalities covering micro-defects in needle tip processing.

[0078] After completing bidirectional reasoning and occlusion determination, based on the temporal anomaly state sequence output by the network, the temporal propagation entropy is calculated to further filter anomaly nodes. The temporal propagation entropy is used to determine the cross-slice propagation effectiveness of node anomalies. The anomaly state is binarized into discrete states of 0 (normal) and 1 (abnormal). The temporal propagation entropy between each substantially anomaly node and its subsequent slice nodes is calculated as cross-validation of the Granger causality test results in step S3. If a node passes the causality test in S3 but its propagation entropy value is lower than 0.1, the causality strength of the node is downgraded to "weak causality" and noted in the final root cause conclusion, but it is not directly removed. The formula for calculating the temporal propagation entropy is: ; in, The temporal transfer entropy from preceding slice node X to subsequent slice node Y. This represents the abnormal state at time t+1 in subsequent slices. This represents the abnormal state at time t in subsequent slices. This represents the abnormal state of the preceding slice at time t. Let represent the probability that the anomalous states of the preceding slice at time t, the subsequent slice at time t, and the subsequent slice at time t+1 occur simultaneously. Given the anomalous states of the preceding slice at time t and the subsequent slice at time t, let be the probability that the subsequent slice is in the corresponding anomalous state at time t+1. Given only the known anomalous state of the subsequent slice at time t, the probability that the subsequent slice will be in the corresponding anomalous state at time t+1.

[0079] For valid anomalous nodes verified by propagation entropy, the SHAP value is used to calculate the coupling contribution of each node, dismantling the linkage amplification mechanism of complex root causes. When calculating the SHAP contribution, an improved dynamic Bayesian network is used as the baseline model, and the posterior probability of the anomalous state in the final inspection slice is used as the model output. For any subset S of nodes, the observed data of its contained nodes are set to the actual batch values, and the data of the remaining nodes are set to the baseline values ​​(i.e., the average parameter values ​​of this process under normal production conditions over the past three months). These values ​​are then input into the network to complete inference, obtaining the final inspection anomaly probability of the corresponding subset. The formula for calculating the SHAP value is: ; in, Let be the SHAP contribution value of the i-th anomalous node, and N be the total number of anomalous nodes after filtering. Let i be a subset of nodes that does not contain the i-th node. This represents the number of nodes in the subset. The joint contribution value of the subset after adding the i-th node. The subset independent contribution value is the final detection anomaly probability of the improved dynamic Bayesian network inference output, which characterizes the degree of contribution of the node subset S to the final detection failure.

[0080] The threshold for the absolute value of SHAP contribution is set to 0.1. Valid anomalous nodes with an absolute value of contribution greater than or equal to 0.1 are identified as core root cause nodes. If the absolute value of SHAP contribution of all substantive anomalous nodes is less than 0.1, it is identified as a multi-factor coupling anomaly. All substantive anomalous nodes are retained, and the causal strength of each node is used as a substitute for contribution ranking. The root cause conclusion is output for process engineers to refer to and triggers the manual review process.

[0081] The coupling relationship type is determined based on the temporal sequence and interaction contribution of the core root cause nodes. The quantitative determination rules for the three types of coupling relationships are as follows: if the SHAP contribution of the preceding node is ≥0.1, and the transmission coefficient between the preceding and following nodes is ≥0.3, and the temporal propagation entropy is ≥0.1, the interaction relationship is determined to be a sequential dependency coupling; when the SHAP contribution of two or more nodes is ≥0.08, and the coupling amplification transfer factor is ≥0.1, the coupling relationship is determined to be a sequential dependency coupling. A SHAP contribution >1.5 indicates amplified dependency coupling; the SHAP contribution of the current order node is <0.08, the SHAP contribution of the subsequent order node is ≥0.1, and the abnormal occlusion probability is high. If the threshold is >0.3 (if the preceding node is a noise node marked by S3, the threshold is relaxed to 0.5), it is determined to be occlusion dependency coupling. If the coupling relationship between nodes satisfies multiple judgment conditions at the same time, the judgment is made according to the priority of occlusion dependency coupling > amplification dependency coupling > sequential dependency coupling. The coupling relationship with the highest priority is output as the main coupling type, and the rest are noted in the conclusion as auxiliary coupling types.

[0082] If the improved dynamic Bayesian network fails to converge after 50 iterations, or if the entropy value of the inference result is greater than 0.8 (indicating high uncertainty of the result), the network inference stage is skipped, and the two nodes with the highest causal strength output by S3 are directly used as the core root cause nodes, marked as "model inference abnormal based on causal strength judgment", and the manual review process is triggered.

[0083] Based on the above analysis results, a hierarchical root cause conclusion is formed, which includes three core levels. The first level is the final inspection explicit failure layer, which corresponds to the final inspection failure mode and clarifies the core failure manifestations of the current batch of probe devices, such as contact impedance fluctuations and coplanarity anomalies. The second level is the process transmission layer, which corresponds to the process transmission compression path and the transferable substantial anomaly nodes, clarifying the complete time sequence path of the anomaly along the process transmission compression path, from the transferable substantial anomaly nodes to the final inspection. The third level is the root cause triggering layer, which corresponds to the three types of node anomaly coupling relationships identified by the improved dynamic Bayesian network: sequential dependency coupling, amplified dependency coupling, and masking dependency coupling. It clarifies that yield anomalies are triggered by the instability of a single core root cause node or by the superposition and amplification of the three types of node anomaly coupling relationships of multiple core root cause nodes.

[0084] The final inspection failure modes, process transmission compression paths, core root cause nodes and SHAP contribution, node anomaly coupling relationship types and quantitative indicators, analysis timestamps and batch numbers analyzed in this study are all organized into reusable anomaly transmission knowledge units. These knowledge units are stored in the probe production anomaly knowledge base in a structured form. When similar final inspection failure modes occur in the future, the system will prioritize the retrieval of similar knowledge units for rapid comparison, directly lock the priority investigation direction and root cause judgment rules, continuously improve the intelligent mining efficiency and judgment accuracy of probe production yield anomalies, and finally output the root cause conclusions of probe production yield anomalies, including single-node instability type or multi-node coupling type.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A probe device yield abnormality root cause intelligent mining method, characterized by: Includes the following steps: Step S1: Obtain probe final inspection data and historical yield abnormal case data. Preprocess the probe final inspection data to obtain failure feature data. Compare the failure feature data with a preset threshold to obtain the final inspection failure mode. Based on the final inspection failure mode, perform association rule mining on the historical yield abnormal case data to obtain the associated sensitive process set. Step S2: Obtain the process production time sequence data of the probe device, construct the process time sequence dependency chain based on the associated sensitive process set and process production time sequence data, and filter nodes and remove branches in the process time sequence dependency chain through the reverse depth-first search algorithm to obtain the process transmission compression path; Step S3: Use the improved Granger causality test to determine the abnormal transmission of the process transmission compression path and screen out the set of substantive abnormal nodes; Step S4: Obtain time-series slice data, input the actual abnormal node set and time-series slice data into the improved dynamic Bayesian network model, output the core root cause node and node abnormal coupling relationship, and generate the root cause conclusion of abnormal production yield of probe device based on the core root cause node and node abnormal coupling relationship. The specific steps of the improved Granger causality test include: For each pair of adjacent process nodes in the process transmission compression path, an autoregressive moving average generalized autoregressive conditional heteroscedasticity model containing the parameters of the preceding process is established as the full model, and a restricted model is obtained by removing the relevant terms of the preceding process parameters from the mean equation of the full model. The full model and the restricted model use the same variance equation structure; Solve for the parameters of the full model and the restricted model, and determine whether there is a Granger causal relationship between the anomalies of the parameters in the preceding process and the anomalies of the parameters in the following process based on the likelihood ratio test; Traverse all adjacent process node pairs in the process transmission compression path, retain process nodes with Granger causality, and remove local noise nodes without Granger causality to obtain the set of substantial abnormal nodes. The improved dynamic Bayesian network model specifically includes: Obtain the set of substantive anomaly nodes, divide the time sequence into time slices according to the continuous time sequence of the probe production process, and obtain the time slice data corresponding to each substantive anomaly node; Constructing an improved dynamic Bayesian network; The improved dynamic Bayesian network consists of multiple temporal slice Bayesian subnetworks; Nodes in adjacent time-series slices are connected by directed edges for anomalous state transitions, and the subnetworks corresponding to different time-series slices are allowed to have different state transition probabilities. Embedding a coupling amplification transition factor in the calculation of state transition probabilities of dynamic Bayesian networks; The coupling amplification transfer factor is calculated based on the total number of time slices with consecutive anomalies and the standardized anomaly deviation value of the actual anomaly nodes within each time slice. The state transition probability is corrected based on the aforementioned coupling amplification transfer factor; The specific method for correcting the state transition probability is as follows: when the previous time slice is in a normal state, the original state transition probability is used; when the previous time slice is in an abnormal state, the coupling amplification transition factor is used to weight and normalize the original state transition probability. An improved dynamic Bayesian network based on modified state transition probabilities performs forward and backward inference to obtain the joint probability of anomalous states in time slices. The abnormal occlusion probability is calculated based on the conditional probability distribution of the improved dynamic Bayesian network, and the abnormal occlusion probability is compared with the occlusion probability threshold to determine whether there is occlusion dependency coupling. The temporal propagation entropy is calculated based on the joint probability of abnormal states in each temporal slice, and nodes with propagation entropy lower than the propagation entropy threshold are removed. Calculate the SHAP contribution value for the remaining nodes after removal, and determine the nodes with an absolute contribution value greater than the contribution threshold as core root cause nodes; The type of coupling relationship is determined based on the temporal sequence and contribution interaction of the core root cause nodes; The final output yields the root cause conclusions of abnormal production yield of the probe device; The root cause conclusions of abnormal production yield of the probe device include the final inspection failure mode, process transmission path, and the type of coupling relationship.

2. The probe device yield abnormality root cause intelligent mining method according to claim 1, wherein: The process of acquiring probe final inspection data and historical yield anomaly case data, preprocessing the probe final inspection data to obtain failure feature data, and comparing the failure feature data with a preset threshold to obtain the final inspection failure mode includes the following specific steps: Acquire probe final inspection data and historical yield anomaly case data; The probe final inspection data is preprocessed to obtain a failure feature dataset; The failure feature dataset is compared with a preset threshold, and the final failure mode is output. Based on the final inspection failure mode, association rule mining is performed on the historical yield anomaly case data to obtain the associated sensitive process set.

3. The probe device yield abnormality root cause intelligent mining method according to claim 2, characterized by: The step of performing association rule mining on historical yield anomaly case data based on final inspection failure modes to obtain a set of associated sensitive processes includes: Based on the final inspection failure modes and historical yield anomaly case data, a transaction set is constructed; Generate a candidate set based on the transaction set and calculate the support. Candidate itemsets are filtered based on support to obtain frequent itemsets; Calculate the confidence and lift of each combination in the frequent itemset; Filter combinations that satisfy the confidence threshold and the lift threshold, and extract the processes corresponding to the combinations to form a set of associated sensitive processes.

4. The intelligent method for identifying the root causes of abnormal production yield in a probe device according to claim 3, characterized in that: The process production time sequence data of the probe device is acquired, and a process time sequence dependency chain is constructed based on the associated sensitive process set and the process production time sequence data, including the following steps: Acquire process production timing data of the probe device, wherein the process production timing data includes process execution sequence data and inter-process influence interface data; Basic nodes are extracted from the set of related sensitive processes. Intermediate processes between basic nodes are automatically supplemented as transmission relay nodes according to the fixed process sequence of probe production to form a complete initial node sequence. The continuous adjacent processes in the initial node sequence are used as nodes, the actual execution order of the processes is used as the direction of the directed edges, and the transmission coefficient of the interface between processes is used as the weight of the directed edges to construct a process time dependency chain in the form of a directed acyclic graph. The propagation coefficient of the directed edge weights is calculated using the maximum information coefficient.

5. The intelligent method for identifying the root causes of abnormal production yield in a probe device according to claim 4, characterized in that: The process propagation compression path is obtained by filtering nodes and eliminating branches in the process time dependency chain using a reverse depth-first search algorithm, as detailed below: Calculate the mutual information value between the parameter time-series data and the final inspection failure feature data of each process node in the process time-series dependency chain, and remove process nodes with mutual information values ​​lower than the mutual information threshold to obtain highly correlated process nodes; A reverse depth-first search algorithm is used to traverse and prune the directed acyclic graph composed of highly correlated process nodes. Cut off directed edges and corresponding branches with a conduction coefficient lower than the conduction coefficient screening threshold, remove isolated nodes and branches without continuous conduction association, retain continuous conduction paths where the conduction coefficient between all adjacent nodes is not lower than the conduction coefficient threshold and the mutual information value of each node is not lower than the mutual information threshold, and generate process conduction compression paths.

6. The intelligent method for identifying the root causes of abnormal production yield in a probe device according to claim 5, characterized in that: The step of using the improved Granger causality test to determine abnormal transmission in the process transmission compression path and screening out the set of substantially abnormal nodes includes: Obtain the compression path of the process transmission and the process production timing data of each process; The stationarity test and conditional heteroscedasticity test were performed sequentially on the production time sequence data of each process. For non-stationary time series data, a linear trend term is first fitted. If the absolute value of the slope of the trend term exceeds a preset threshold, the trend term is retained as an independent outlier. If the sequence is still non-stationary after removing significant trend terms, a first-order difference transformation is performed. For time series data exhibiting conditional heteroscedasticity, an improved Granger causality test method based on the autoregressive moving average generalized autoregressive conditional heteroscedasticity model is used to determine anomaly propagation.

Images