A horizontal copper production full-process tracing method and system

Abstract

The application discloses a horizontal copper deposition production whole-process tracing method and system, and relates to the technical field of data acquisition and management. In the method, first, a characteristic vector of cross-process parameter correlation data is acquired, and a parameter influence intensity index is calculated based on a joint probability distribution of drilling and deslagging components; if the index exceeds a threshold value, an abnormal point is screened by calculating a distance matrix of hole wall roughness and historical samples to obtain an abnormal parameter set; then, a copper deposition defect process node is extracted, a deviation transmission gain is calculated to determine a root cause positioning sequence; when the sequence meets a similarity standard with a historical sequence, the sequence is fused to an initial atlas to obtain a target tracing atlas; finally, real-time data is input into the atlas, a defect tracing path is generated based on a highest connected probability, and a parameter adjustment instruction sequence is generated accordingly. The technical scheme provided by the application can improve the accuracy of horizontal copper deposition production whole-process tracing.

Description

Technical Field

[0001] This invention belongs to the field of product data acquisition and management technology, specifically a method and system for tracing the entire process of horizontal copper plating production. Background Technology

[0002] With the rapid development of printed circuit board manufacturing technology, the chemical copper plating process, as a key step in achieving hole wall metallization, faces increasingly stringent requirements for product quality control and reliability. End-to-end quality traceability, a crucial aspect of production management, enables real-time tracking of parameters at each stage, from drilling and desmearing to activation and copper plating, preventing quality risks arising from accumulated process deviations. To meet the needs of different product specifications, process routes, and production batches, quality traceability technology is exhibiting a trend towards refinement. Its accuracy and reliability in complex process coupling environments are of paramount importance for ensuring product quality and production efficiency.

[0003] Currently, the quality traceability method in horizontal copper plating production mainly adopts single-process parameter monitoring technology. This involves setting independent detection points at each process stage to record process parameters and comparing the parameter values ​​with preset specification ranges to determine whether the process is functioning correctly. This parameter range-based detection method determines the process status by comparing the parameter values ​​collected from each process with preset judgment thresholds. It can identify basic parameter deviations in individual processes and has been widely used in the field of PCB manufacturing quality management.

[0004] However, due to the complexity of the production process and the variability of inter-process coupling relationships, the formation mechanism of quality defects is becoming increasingly complex and varied. In practical applications, traceability methods based on independent monitoring of a single process are unable to suppress the interference of factors such as cross-process parameter interactions on the traceability results, and the accuracy of defect root cause location is difficult to guarantee. In particular, when small deviations in upstream processes are transmitted and accumulated through multiple processes, due to the implicit correlation between parameters in each process, existing methods that use simple single-point parameter judgments are prone to deviations or even misidentification in root cause location, thereby reducing the accuracy of traceability throughout the entire horizontal copper plating production process. Summary of the Invention

[0005] To address the above problems, this invention provides a method and system for tracing the entire horizontal copper plating production process. This method solves the problem that existing traceability methods cannot guarantee the accuracy of defect root cause location under the influence of cross-process parameter interactions, and can improve the accuracy of tracing the entire horizontal copper plating production process.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Obtain the feature vectors corresponding to cross-process parameter correlation data; Based on the feature vector, borehole parameter components and desizing effect components are extracted, the joint probability distribution between the borehole parameter components and desizing effect components is calculated, and the parameter influence intensity index is determined based on the joint probability distribution. If the influence of the parameter on the intensity index exceeds a preset threshold, then the distance matrix between the hole wall roughness feature in the feature vector and the historical working condition sample is calculated, and potential abnormal points deviating from the standard working condition are screened out based on the distance matrix to obtain an abnormal parameter set. Extract the process nodes corresponding to the copper plating defects from the abnormal parameter set, calculate the deviation propagation gain between adjacent process nodes according to the process sequence, construct the influence chain based on the deviation propagation gain, and determine the root cause localization sequence. Calculate the similarity between the root cause location sequence and the historical traceability sequence in the historical production log data. When the similarity meets a preset similarity threshold, merge the node connection relationship in the root cause location sequence into the initial parameter relationship graph to obtain the target traceability graph. The real-time monitoring data is input into the target traceability map, the connectivity probability between each node in the target traceability map is calculated, a quality defect backtracking path is generated based on the path corresponding to the highest connectivity probability, and an upstream impact correlation index is extracted based on the quality defect backtracking path to generate a parameter adjustment instruction sequence.

[0007] By adopting the above technical solution and obtaining the feature vectors corresponding to the cross-process parameter correlation data, the limitations of existing single-process independent monitoring are overcome, providing a data foundation for cross-process parameter interaction analysis. Based on this, by extracting drilling parameter components and degumming effect components and calculating their joint probability distribution, the parameter influence intensity index is determined, achieving a quantitative assessment of the inter-process correlation influence. This makes the originally implicit cross-process parameter interaction relationship explicit, avoiding the problem of existing methods that isolate process parameters and ignore their correlation. When the parameter influence intensity index exceeds a preset threshold, the distance matrix between the borehole wall roughness feature and historical working condition samples is calculated to screen potential anomalies and obtain an abnormal parameter set. Compared with the simple single-point parameter threshold judgment of existing methods, this can identify abnormal patterns deviating from standard working conditions in a multi-dimensional parameter space, improving the accuracy of anomaly identification. Furthermore, by extracting process nodes from the abnormal parameter set and calculating the deviation transmission gain between adjacent process nodes to construct an influence chain and determine the root cause localization sequence, the quantitative tracking of how small deviations in upstream processes are transmitted and accumulated through multiple processes is achieved, solving the problem that existing methods struggle to trace the deviation transmission path. By matching the root cause localization sequence with historical tracing sequences based on similarity and fusing them into a parameter relationship graph to obtain the target tracing graph, a systematic learning and knowledge reuse of historical defect cases is achieved, enabling continuous optimization and evolution of the tracing graph. Finally, by inputting real-time monitoring data into the target tracing graph and generating quality defect backtracking paths based on the connectivity probability between nodes, the true root cause process and key influencing parameters can be quickly and accurately located. This avoids the root cause localization deviations or identification errors caused by the inability of existing methods to accurately identify the interaction effects of cross-process parameters, thereby improving the accuracy of the entire process traceability in horizontal copper plating production.

[0008] Optionally, cross-process parameter correlation data for each process in the pre-production stage is obtained, including measured values ​​of process parameters and measured values ​​of quality characteristics; the normalized deviation of the measured values ​​of process parameters is calculated based on the process specification settings of each process; the normalized deviation is concatenated with the corresponding measured values ​​of quality characteristics to generate the feature vector.

[0009] Optionally, the drilling parameter component representing the drilling speed and feed rate, and the descaling effect component representing the residual adhesive thickness are separated from the feature vector; the co-occurrence frequency of the drilling parameter component and the descaling effect component within a preset value range is statistically analyzed, and the co-occurrence frequency is converted into a joint probability distribution; the independent occurrence frequency of the drilling parameter component and the descaling effect component within their respective component value ranges is statistically analyzed, and the independent occurrence frequency is converted into a corresponding marginal probability distribution; the marginal probability distributions of the drilling parameter component and the descaling effect component are multiplied to generate an independent probability distribution; the relative entropy between the joint probability distribution and the independent probability distribution is calculated, and the relative entropy is used as an index of the parameter influence intensity.

[0010] Optionally, if the influence of the parameter on the intensity index exceeds a preset threshold, then the hole wall roughness feature in the feature vector is extracted; a reference domain containing multiple historical working condition samples is obtained; the Mahalanobis distance between the hole wall roughness feature and each of the historical working condition samples in the reference domain is calculated to form the distance matrix; sample points with distance values ​​greater than a preset distance judgment threshold are extracted from the distance matrix as potential anomaly points, and the parameters corresponding to the potential anomaly points are summarized into the anomaly parameter set.

[0011] Optionally, the covariance matrix of the historical working condition samples in the reference domain is calculated, and the inverse matrix of the covariance matrix is ​​solved; using the inverse matrix as weights, the quadratic difference value between the hole wall roughness feature and each of the historical working condition samples is calculated; the quadratic difference value is used as the Mahalanobis distance to fill the corresponding element position of the matrix to generate the distance matrix.

[0012] Optionally, the abnormal parameter set is traversed, and target parameters matching the copper bonding strength defect are selected from the abnormal parameter set according to a preset defect parameter mapping table. The process in which the target parameter is located is taken as the process node. The parameter change and quality characteristic change between adjacent process nodes are obtained. The ratio of the quality characteristic change to the parameter change is taken as the deviation propagation gain. The corresponding process nodes are connected in series according to the increasing order of the deviation propagation gain to generate the influence chain, and the node sequence in the influence chain is taken as the root cause localization sequence.

[0013] Optionally, the feature parameters of each process node in the root source location sequence and the historical tracing sequence are extracted; the spatial distance between the feature parameters of each process node in the root source location sequence and the historical tracing sequence is calculated to construct a local cost matrix; based on the local cost matrix, the cumulative cost of all aligned paths from the start point to the end point of the sequence is calculated, and the minimum cumulative cost is taken as the sequence normalization distance; the reciprocal of the sum of the sequence normalization distance and the preset adjustment coefficient is calculated, and the reciprocal is taken as the similarity; when the similarity is greater than the preset similarity threshold, the newly added node connection relationship in the root source location sequence is extracted, and the newly added node connection relationship is added as a directed edge to the initial parameter relationship graph to obtain the target tracing graph.

[0014] Optionally, real-time monitoring data containing the current batch process deviation is mapped to the corresponding node of the target traceability graph, and the target defect node corresponding to the current copper plating defect is located; based on the directed edge connection relationship in the target traceability graph, all connected branches pointing from each upstream process node to the target defect node are traversed, and the extracted connected node sequence is used as a candidate backtracking path; combined with the historical confidence of each directed edge in the optimized traceability graph, the conditional probability of the real-time monitoring data being transmitted between adjacent nodes of the candidate backtracking path is calculated; the conditional probabilities between all adjacent nodes on each candidate backtracking path are multiplied to obtain the connectivity probability of the candidate backtracking path; the connectivity probabilities of each candidate backtracking path are compared, and the candidate backtracking path corresponding to the highest connectivity probability is selected as the quality defect backtracking path.

[0015] Optionally, the parameter deviation and copper plating constraint margin of the main root cause process are extracted from the quality defect backtracking path as the upstream influence correlation index; it is determined whether the copper plating constraint margin is lower than the preset acceptable lower limit; if the copper plating constraint margin is lower than the preset acceptable lower limit, the influence gradient of the parameter deviation on the copper plating constraint margin is calculated; the adjustment range for compensating the parameter deviation is calculated according to the influence gradient, the main root cause process is determined as the target adjustment node, and the identification information of the target adjustment node is combined with the adjustment range to generate the parameter adjustment instruction sequence.

[0016] Secondly, embodiments of this application provide a horizontal copper plating production process traceability system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the horizontal copper plating production process traceability system to perform the method described in the first aspect and any possible implementation thereof.

[0017] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By adopting the above technical solution and obtaining the feature vectors corresponding to the cross-process parameter correlation data, the limitations of existing single-process independent monitoring are overcome, providing a data foundation for cross-process parameter interaction analysis. Based on this, by extracting drilling parameter components and degumming effect components and calculating their joint probability distribution, the parameter influence intensity index is determined, achieving a quantitative assessment of the inter-process correlation influence. This makes the originally implicit cross-process parameter interaction relationship explicit, avoiding the problem of existing methods that isolate process parameters and ignore their correlation. When the parameter influence intensity index exceeds a preset threshold, the distance matrix between the borehole wall roughness feature and historical working condition samples is calculated to screen potential anomalies and obtain an abnormal parameter set. Compared with the simple single-point parameter threshold judgment of existing methods, this can identify abnormal patterns deviating from standard working conditions in a multi-dimensional parameter space, improving the accuracy of anomaly identification. Furthermore, by extracting process nodes from the abnormal parameter set and calculating the deviation transmission gain between adjacent process nodes to construct an influence chain and determine the root cause localization sequence, the quantitative tracking of how small deviations in upstream processes are transmitted and accumulated through multiple processes is achieved, solving the problem that existing methods struggle to trace the deviation transmission path. By matching the root cause localization sequence with historical tracing sequences based on similarity and fusing them into a parameter relationship graph to obtain the target tracing graph, a systematic learning and knowledge reuse of historical defect cases is achieved, enabling continuous optimization and evolution of the tracing graph. Finally, by inputting real-time monitoring data into the target tracing graph and generating quality defect backtracking paths based on the connectivity probability between nodes, the true root cause process and key influencing parameters can be quickly and accurately located. This avoids the root cause localization deviations or identification errors caused by the inability of existing methods to accurately identify the interaction effects of cross-process parameters, thereby improving the accuracy of the entire process traceability in horizontal copper plating production. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for tracing the entire production process of horizontal copper plating disclosed in an embodiment of this application. Figure 2 This is another schematic diagram of a method for tracing the entire production process of horizontal copper plating disclosed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a system provided in an embodiment of this application.

[0019] In the diagram: 301, Central Processing Unit; 302, Read-Only Memory; 303, Random Access Memory; 304, Bus; 305, Input / Output Interface; 306, Input Section; 307, Output Section; 308, Storage Section; 309, Communication Section; 310, Driver; 311, Removable Media. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solution, the present invention will be described in detail below with reference to embodiments. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.

[0021] It should be noted that, in this document, the terms "comprising," "including," and any other variations 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. Specific examples have been used in this document to illustrate the principles and implementation methods of the present invention. These examples are merely for the purpose of helping to understand the method and core ideas of the present invention. The above are only preferred embodiments of the present invention. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or variations without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of the present invention to other situations without modification, should all be considered within the scope of protection of the present invention.

[0022] This application provides a method for tracing the entire production process of horizontal copper plating, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a method for tracing the entire production process of horizontal copper plating according to an embodiment of this application. The method is applied to a system, which refers to a hardware and software integrated platform capable of executing a traceability program for the entire production process of horizontal copper plating. The system can execute a traceability program for the entire production process of horizontal copper plating. The method includes steps 101 to 106, as follows: Step 101: Obtain the feature vector corresponding to the cross-process parameter association data.

[0023] In this embodiment, cross-process parameter correlation data refers to the process parameters and quality characteristic data collected during the horizontal copper plating process, spanning multiple processes such as drilling, desmearing, activation, and copper plating. This data reflects the correlation and mutual influence of parameters between different processes. Examples include parameters such as drill bit speed and feed rate in the drilling process, chemical concentration and temperature in the desmearing process, and soaking time in the activation process, as well as quality characteristics measured after each process, such as hole wall roughness and residual adhesive thickness. A feature vector represents a multi-dimensional numerical vector transformed from cross-process parameter correlation data through mathematical processing. Each dimension represents a specific process parameter deviation or quality characteristic value. For example, a feature vector may contain values ​​for multiple dimensions such as [drill bit speed deviation, feed rate deviation, residual adhesive thickness, hole wall roughness, etc.].

[0024] Specifically, the process first obtains parameter data for each step of the pre-production stage from the production system. This includes measured values ​​of process parameters for steps such as drilling, desmearing, activation, and copper plating, such as drill bit speed, feed rate, desmearing temperature, and solution concentration, as well as measured values ​​of quality characteristics, such as hole wall roughness, residual adhesive thickness, and copper layer thickness. Then, using the process specification settings for each step as a benchmark, the normalized deviation of the measured process parameters is calculated as: Parameter Deviation = (Measured Value - Set Value) / Set Value. Finally, the calculated normalized deviation is concatenated with the corresponding measured values ​​of quality characteristics to generate a multi-dimensional feature vector. This vector contains deviation information for multiple process parameters and quality characteristic information, providing a data foundation for subsequent analysis.

[0025] In one possible implementation, obtaining the feature vector corresponding to the cross-process parameter correlation data specifically includes steps 1011-1013, as follows: Step 1011: Obtain cross-process parameter correlation data for each process in the pre-production stage. The cross-process parameter correlation data includes measured values ​​of process parameters and measured values ​​of quality characteristics.

[0026] Specifically, the data collection scope for the pre-production stage is first determined, including all control parameters and quality inspection points for key processes such as drilling, desmearing, activation, and copper plating. Then, real-time measured values ​​of process parameters are collected from various sensors, detection equipment, and control systems on the production line. These parameters include drill bit speed and feed rate in the drilling process; solution concentration, temperature, and stirring speed in the desmearing process; activation time and pH value of the activation solution in the activation process; and copper concentration in the copper plating process. Simultaneously, after each process is completed, measured values ​​of quality characteristics are collected using appropriate detection methods, including hole wall roughness after drilling (e.g., 1.2 μm), hole diameter accuracy (e.g., ±0.05 mm), residual adhesive thickness after desmearing (e.g., 0.2 μm), activated layer thickness after activation (e.g., 0.05 μm), copper layer thickness after copper plating (e.g., 25 μm), and copper layer adhesion (e.g., 0.8 N / mm), etc. Next, the collected measured values ​​of process parameters and quality characteristics are correlated according to timestamps and workpiece batch numbers to form a cross-process parameter correlation data table containing complete process information. Finally, preliminary outlier detection and missing value processing are performed on the collected data to ensure data integrity and accuracy, providing a reliable data foundation for subsequent analysis.

[0027] Step 1012: Based on the process specification settings of each process, calculate the normalized deviation of the measured values ​​of the process parameters.

[0028] Specifically, the process specifications for each step are first obtained. These specifications are standard parameters defined in the production process documents, such as the standard drill bit speed and standard feed rate for drilling, and the standard solution concentration and standard temperature for descaling. Then, for each process parameter, the difference between the measured value and the specified value is calculated, i.e., the measured value of the process parameter is subtracted from the corresponding process specification setting. Next, the calculated difference is divided by the process specification setting to obtain the relative deviation value. For example, if the measured drill bit speed is 20,000 rpm, while the specified setting is 19,000 rpm, the normalized deviation is (20,000 - 19,000) / 19,000 ≈ 0.0526, indicating that the speed is about 5.26% too high. Similarly, if the measured descaling solution concentration is 4.5%, while the specified setting is 4.8%, the normalized deviation is (4.5 - 4.8) / 4.8 ≈ -0.0625, indicating that the concentration is about 6.25% too low. Finally, the normalized deviations of all process parameters were organized into a unified data structure for subsequent concatenation with measured values ​​of quality characteristics. Through normalization, parameter deviations of different types and magnitudes were converted into relative percentages, facilitating subsequent analysis and comparison on a unified scale.

[0029] Step 1013: Concatenate the normalized deviation with the corresponding measured quality feature values ​​to generate a feature vector.

[0030] Specifically, the dimensions and structure of the feature vector are first determined, typically by arranging the parameters and quality characteristics according to the process sequence. For example, the feature vector can be designed as: [Drilling process parameter deviation, Drilling quality characteristic, Descaling process parameter deviation, Descaling quality characteristic, Activation process parameter deviation, Activation quality characteristic, Copper plating process parameter deviation, Copper plating quality characteristic]. Then, the calculated normalized deviations of each process parameter are filled into the corresponding dimensions of the feature vector according to predetermined positions. For example, a drill bit speed deviation of 0.0526 is filled into the first position of the feature vector, a feed rate deviation of -0.0333 is filled into the second position, and so on. Next, the measured values ​​of the quality characteristics of each process are also filled into the corresponding positions of the feature vector. For example, a hole wall roughness of 1.2 μm is filled into the position of the drilling quality characteristic in the feature vector, a residual adhesive thickness of 0.2 μm is filled into the position of the descaling quality characteristic, and so on. For quality characteristic data, measured values ​​can be used directly, or their deviations from standard values ​​can be calculated first and then filled in, depending on the needs of subsequent analysis. Finally, check that the generated feature vector is complete, ensuring that all key parameters and features are included. After generating the feature vector, it will serve as input data for subsequent analysis steps, used to extract borehole parameter components and desizing effect components, calculate the joint probability distribution, and perform other operations.

[0031] Step 102: Extract the borehole parameter components and the degumming effect components based on the feature vector, calculate the joint probability distribution between the borehole parameter components and the degumming effect components, and determine the parameter influence intensity index based on the joint probability distribution.

[0032] In this embodiment, the drilling parameter component refers to the set of values ​​in the feature vector that characterize the key parameters of the drilling process, mainly including drilling speed deviation and feed rate deviation. These parameters directly affect the quality of PCB board drilling and the hole wall condition. For example, excessively high drill speed may cause the hole wall to overheat, while excessively high feed rate may cause the hole wall to tear. The adhesive removal effect component represents the set of values ​​in the feature vector that reflect the effect of the adhesive removal process, mainly including residual adhesive thickness data. Residual adhesive thickness is a key indicator for measuring the effect of the adhesive removal process, and it directly affects the bonding strength between the copper layer and the substrate in the subsequent copper plating process. The joint probability distribution refers to the probability distribution function of the co-occurrence of the drilling parameter component and the adhesive removal effect component, used to quantify the statistical correlation between the two. The parameter influence intensity index represents the quantitative value of the degree of influence of drilling parameters on the adhesive removal effect, usually measured by relative entropy and KL divergence.

[0033] Specifically, firstly, the drilling parameter components, representing drilling speed and feed rate, and the descaling effect component, representing residual adhesive thickness, are separated from the feature vector. Then, the co-occurrence frequency of these two components within a preset numerical range is calculated. The value range of the drilling parameter components is divided into multiple intervals, such as speed deviation [-5%, -2%], [-2%, 0%], [0%, 2%], [2%, 5%]. Similarly, the descaling effect component is also divided into multiple intervals, such as residual adhesive thickness [0.1μm, 0.5μm], [0.5μm, 1.0μm], [1.0μm, 1.5μm]. By counting the frequency of each pair of interval combinations and dividing by the total number of samples, the co-occurrence frequency is transformed into a joint probability distribution P(X, Y). Next, the independent occurrence frequencies of the drilling parameter components and the descaling effect component within their respective intervals are calculated and transformed into marginal probability distributions P(X) and P(Y). Finally, these two marginal probability distributions are multiplied to obtain the independent probability distribution P(X)×P(Y). Finally, the relative entropy between the joint probability distribution and the independent probability distribution is calculated, and this relative entropy value is used as an index of the parameter influence intensity. This index quantifies the degree of interdependence between borehole parameters and the degumming effect.

[0034] In one possible implementation, borehole parameter components and desizing effect components are extracted based on feature vectors. The joint probability distribution between the borehole parameter components and the desizing effect components is calculated. The parameter influence intensity index is determined based on the joint probability distribution. Specifically, steps 1021-1023 are included, as follows: Step 1021: Separate the drilling parameter components representing the drilling speed and feed rate, and the adhesive removal effect component representing the residual adhesive thickness from the feature vector; count the co-occurrence frequency of the drilling parameter components and the adhesive removal effect component in the preset value range, and convert the co-occurrence frequency into a joint probability distribution.

[0035] In this embodiment, the joint probability distribution refers to a probability model characterizing the co-occurrence of drilling parameter components, drilling speed and feed rate, and desizing effect components, including residual adhesive thickness. It describes the probabilistic relationship between specific drilling parameter values ​​and specific desizing effect values ​​within a preset numerical range. For example, it describes the probability that the residual adhesive thickness falls within a specific range when the drilling speed and feed rate are both within a certain range. The joint probability distribution reflects the statistical dependence between process parameters and quality characteristics and can be used to assess the degree of interaction between different process parameters.

[0036] Specifically, firstly, drilling parameter components representing drilling speed and feed rate are extracted from the previously generated feature vector. These parameter components are typically normalized deviation values ​​related to the drilling process in the feature vector. Simultaneously, the degumming effect component representing residual adhesive thickness is extracted; this is a value in the feature vector reflecting the quality characteristics of the degumming effect. Then, the value range of the drilling parameter components is divided into several discrete intervals. For example, the drilling speed deviation can be divided into intervals such as [-0.1, -0.05), [-0.05, 0), [0, 0.05), and [0.05, 0.1), and the feed rate deviation is similarly divided. Similarly, the value range of the degumming effect component and residual adhesive thickness is also divided into several discrete intervals, such as [0.1, 0.2), [0.2, 0.3), and [0.3, 0.4) μm. Next, all collected sample data are traversed, and the frequency of each combination of drilling parameter intervals and each degumming effect interval is counted to form a frequency counting matrix. For example, the number of samples with residual adhesive thickness in the range of [0.2, 0.3) μm is counted under the conditions that the drilling speed deviation is in the range of [-0.05, 0) and the feed rate deviation is in the range of [0, 0.05). Finally, each element in the frequency counting matrix is ​​divided by the total number of samples to obtain the corresponding joint probability value, thereby constructing a complete joint probability distribution table or matrix.

[0037] Step 1022: Calculate the independent occurrence frequency of the drilling parameter component and the descaling effect component in the corresponding component value range, and convert the independent occurrence frequency into the corresponding marginal probability distribution.

[0038] In this embodiment, the marginal probability distribution refers to calculating the probability distribution of the drilling parameter component and the desizing effect component separately within their respective value intervals, without considering the interrelationships between them. The marginal probability distribution of the drilling parameter component reflects the probability of the drilling speed and feed rate occurring in each preset interval, while the marginal probability distribution of the desizing effect component represents the probability of the residual adhesive thickness occurring in each preset interval. The marginal probability distribution can be viewed as a projection of the joint probability distribution onto its respective dimensions. It eliminates the correlation information between variables, retains only the statistical characteristics of individual variables, and provides basic data for subsequent calculation of independent probability distributions.

[0039] Specifically, the frequency of occurrence of drilling parameter components, specifically the deviations in drilling speed and feed rate, within each preset value interval is statistically analyzed. The method involves iterating through all sample data, recording the number of samples where the drilling speed deviation falls within each preset interval (e.g., [-0.1, -0.05), [-0.05, 0)), and the number of samples where the feed rate deviation falls within each preset interval. Then, the number of samples in each interval is divided by the total number of samples to obtain the marginal probability value of the drilling parameter component in each interval. Similarly, the same statistical process is performed on the degumming effect component, counting the number of samples where the residual adhesive thickness falls within each preset interval (e.g., [0.1, 0.2), [0.2, 0.3) μm), and then dividing by the total number of samples to obtain the marginal probability distribution of the degumming effect component. Finally, marginal probability distribution tables for the drilling parameter components and the degumming effect component are generated, respectively. These two tables characterize the statistical distribution characteristics of the two types of parameters without considering their mutual influence.

[0040] Step 1023: Multiply the marginal probability distributions of the borehole parameter components and the marginal probability distributions of the degumming effect components to generate independent probability distributions; calculate the relative entropy between the joint probability distribution and the independent probability distributions, and use the relative entropy as an index of the parameter influence intensity.

[0041] In this embodiment, relative entropy, also known as KL divergence, is a statistical indicator that measures the difference between two probability distributions, used to quantify the information difference between a joint probability distribution and independent probability distributions. A larger relative entropy value indicates a more significant difference between the two distributions, and also signifies a stronger statistical correlation between drilling parameters and the deslagging effect. In this application, relative entropy is used as a parameter influence strength indicator to quantify the degree of influence of drilling process parameter adjustments on the deslagging effect. For example, a relative entropy of 0 indicates that the two process parameters are completely independent and unrelated, while a larger relative entropy indicates a more significant influence of drilling parameters on the deslagging effect.

[0042] Specifically, firstly, the marginal probability distributions of the obtained borehole parameter components and the marginal probability distributions of the deslagging effect components are multiplied to generate independent probability distributions. Specifically, for each borehole parameter interval combination i and deslagging effect interval j, the independent probability is calculated as the marginal probability of the borehole parameter in interval i multiplied by the marginal probability of the deslagging effect in interval j. This independent probability distribution represents the probability distribution under the assumption that the two components are completely independent. Then, the relative entropy between the obtained joint probability distribution and the independent probability distribution is calculated. Specifically, for each borehole parameter interval combination i and deslagging effect interval j, the logarithm ratio of the joint probability to the independent probability is calculated, and then a weighted sum is performed using the joint probability as the weight. The relative entropy is equal to the sum of the logarithms of the joint probability multiplied by the logarithm of the independent probability for all interval combinations. Finally, the calculated relative entropy value is used as a parameter influence intensity index to quantify the degree of influence of borehole parameters on the deslagging effect.

[0043] Step 103: If the parameter influence intensity index exceeds the preset threshold, calculate the distance matrix between the hole wall roughness feature in the feature vector and the historical working condition sample, and filter out potential abnormal points that deviate from the standard working condition based on the distance matrix to obtain the abnormal parameter set.

[0044] In this embodiment, the hole wall roughness feature refers to the numerical value in the feature vector that characterizes the surface roughness of the hole wall after PCB drilling. It is an important indicator for evaluating drilling quality. Excessive hole wall roughness can affect the adhesion and uniformity of subsequent copper plating. The historical operating condition sample represents a set of parameter and quality characteristic data recorded under standard operating conditions during past production processes, serving as a reference benchmark for evaluating the current production status. The distance matrix refers to the mathematical distance between the current hole wall roughness feature and each sample in the historical operating condition sample set, typically Mahalanobis distance. This matrix is ​​used to quantify the deviation between the current operating condition and the historical standard operating condition. Potential outliers refer to data points in the distance matrix whose distance from the historical standard operating condition exceeds a preset threshold; these points may represent process anomalies or parameter deviations. The set of anomalous parameters represents the set of process parameters extracted from the potential outliers; these parameters may be key factors causing process deviations.

[0045] Specifically, the process begins by determining whether the calculated parameter influence intensity index exceeds a preset threshold, such as 0.3. If it does, the borehole wall roughness feature data is extracted from the feature vector. Then, a reference domain containing multiple historical operating condition samples is obtained; these samples typically come from high-quality batch data from historical production. Next, the Mahalanobis distance between the current borehole wall roughness feature and each historical operating condition sample in the reference domain is calculated. Specifically, the covariance matrix of the historical operating condition samples in the reference domain is first obtained, and its inverse is calculated. Then, using this inverse matrix as weights, the quadratic difference value between the current borehole wall roughness feature and each historical operating condition sample is calculated; this difference value is the Mahalanobis distance. All calculated Mahalanobis distances are then filled into the corresponding matrix positions to form a complete distance matrix. Finally, sample points with distance values ​​greater than a preset distance threshold, such as 2.5, are extracted from the distance matrix as potential anomalies. The parameters corresponding to these potential anomalies are summarized into an anomaly parameter set to provide data support for subsequent root cause analysis.

[0046] Step 104: Extract the process nodes corresponding to the copper plating defects from the abnormal parameter set, calculate the deviation propagation gain between adjacent process nodes according to the process sequence, construct the influence chain based on the deviation propagation gain, and determine the root cause localization sequence.

[0047] In this embodiment, process nodes refer to various process steps in the horizontal copper plating production process, such as drilling, descaling, activation, and copper plating. Each node has specific process parameters and quality characteristics. Deviation propagation gain represents the amplification or attenuation ratio of parameter deviation propagation between adjacent process nodes, used to quantify the impact of upstream process parameter changes on the quality characteristics of downstream processes. The influence chain refers to a sequence of process nodes linked together according to the deviation propagation relationship, showing the propagation path of defect formation. The root cause localization sequence represents a sequence of process nodes ordered in ascending order of deviation propagation gain, used to determine the root cause process and propagation path of quality defects.

[0048] Specifically, the process begins by traversing the obtained set of abnormal parameters. Based on a pre-established defect-parameter mapping table, target parameters related to copper plating bonding defects, such as residual adhesive thickness and activation time, are selected. The processes containing these target parameters, such as desmearing and activation, are then identified as process nodes. Next, the parameter and quality characteristic changes between adjacent process nodes are obtained from the production data, such as the drill bit speed change in the drilling process and the residual adhesive thickness change in the desmearing process. Then, the ratio of quality characteristic change to parameter change is calculated as the deviation propagation gain, for example, the percentage change in residual adhesive thickness divided by the percentage change in drill bit speed. Subsequently, the corresponding process nodes are linked together in ascending order of the calculated deviation propagation gain to form an influence chain. The larger the gain value of a process node, the more significant its parameter change has on the final quality. Finally, the sequence of nodes in the influence chain is used as the root cause localization sequence. The last node in this sequence, the node with the largest gain, is usually considered the most likely root cause process of the defect.

[0049] In one possible implementation, the process nodes corresponding to the copper plating defects are extracted from the abnormal parameter set, the deviation propagation gain between adjacent process nodes is calculated according to the process sequence, an influence chain is constructed based on the deviation propagation gain, and the root cause localization sequence is determined. Specifically, this includes steps 1041-1043, as follows: Step 1041: Traverse the abnormal parameter set, and according to the preset defect parameter mapping table, select the target parameter that matches the copper bonding strength defect from the abnormal parameter set, and take the process where the target parameter is located as the process node.

[0050] Specifically, the process first acquires the set of abnormal parameters identified in the system. This set includes process parameters that have been detected exceeding normal ranges in each step of the production process. Then, it reads the system's pre-set defect parameter mapping table, which contains the correspondence between copper plating bonding defects and various process parameters. Next, each parameter in the abnormal parameter set is iterated and checked, and compared with the parameter items related to copper plating bonding defects in the defect-parameter mapping table. When an abnormal parameter is found to be associated with copper plating bonding defects in the mapping table, it is marked as a target parameter and added to the target parameter set. For example, if the abnormal parameter set includes parameters such as the temperature parameter of the descaling process exceeding the upper limit or the pH value of the activation process falling below the lower limit, and the mapping table shows that these parameters are related to copper plating bonding defects, then they are selected as target parameters. Finally, based on the process flow stage to which each target parameter belongs, the corresponding process node is determined, such as the descaling process node, the activation process node, or the copper plating process node. These nodes will serve as the basic units for subsequent analysis.

[0051] Step 1042: Obtain the parameter changes and quality characteristic changes between adjacent process nodes; use the ratio of quality characteristic changes to parameter changes as the deviation transfer gain.

[0052] Specifically, first, all identified process nodes are determined and arranged chronologically according to the copper plating production process. Then, for each pair of adjacent process nodes, the difference between the actual and standard values ​​of the target parameters of the preceding process node is obtained, and the parameter change is calculated. Simultaneously, the difference between the actual and standard values ​​of the quality characteristics of the following process node is obtained, and the quality characteristic change is calculated. For example, if the standard value of the temperature parameter for the descaling process is 60℃ and the actual value is 65℃, then the parameter change is 5℃; if the standard value of the bonding strength characteristic for the subsequent activation process is 100 N / cm... 2 The actual value is 90 N / cm. 2 The change in mass characteristic is -10 N / cm. 2 Next, for each pair of adjacent process nodes, the ratio of the change in quality characteristics to the change in parameters is calculated to obtain the deviation propagation gain. In the example above, the deviation propagation gain is -10 ÷ 5 = -2, indicating that for every 1°C increase in the temperature of the desizing agent, the bonding strength of the activation process decreases by 2 N / cm. 2 Finally, the calculated deviation propagation gain values ​​for each pair of adjacent process nodes are stored to prepare for subsequent process node sorting and influence chain construction.

[0053] Step 1043: Connect the corresponding process nodes in series according to the order of increasing deviation propagation gain to generate an influence chain, and use the node sequence in the influence chain as the root cause localization sequence.

[0054] Specifically, first, all adjacent process node pairs and their corresponding deviation propagation gain values ​​are collected and calculated. Then, these node pairs are sorted in ascending order of deviation propagation gain value, with smaller gain values ​​at the beginning and larger gain values ​​at the end. Next, starting with the node pair with the smallest deviation propagation gain, it is used as the starting part of the influence chain. For the second sorted node pair, it is checked whether its preceding node is already in the chain; if so, the following node is added to the end of the chain; otherwise, this pair is treated as a new branch of the chain. This process is repeated for all sorted node pairs to form a complete influence chain. For example, suppose there are node pairs A→B, B→C, and C→D, which are sorted by gain and connected sequentially to form the chain A→B→C→D. Finally, the node sequence in the constructed influence chain, i.e., A, B, C, D, is used as the root cause localization sequence. This sequence indicates that the parameter deviation most likely originates from node A and then propagates sequentially to B, C, and D, with the process corresponding to node D having the most direct and significant impact on the quality characteristics.

[0055] Step 105: Calculate the similarity between the root source location sequence and the historical traceability sequence in the historical production log data. When the similarity meets the preset similarity threshold, merge the node connection relationship in the root source location sequence into the initial parameter relationship graph to obtain the target traceability graph.

[0056] In this embodiment, the historical traceability sequence refers to the verified root cause tracing path of quality defects recorded in historical production log data, reflecting the process impact chain under similar defect situations in the past. Similarity represents the degree of matching between the current root cause localization sequence and the historical traceability sequence, typically calculated using sequence comparison algorithms, such as Dynamic Time Warping (DTW). The initial parameter relationship graph refers to a pre-established network structure representing the correlation between process parameters and quality characteristics. The target traceability graph represents the optimized parameter relationship graph after incorporating newly identified node connection relationships, used for subsequent defect tracing and parameter optimization.

[0057] Specifically, firstly, feature parameters of each process node in the root cause localization sequence and the historical traceability sequence in the historical production log data are extracted, such as the rotation speed and feed rate of the drilling node, and the temperature and concentration of the deslagging node. Then, the spatial distance between the feature parameters of each process node in the current root cause localization sequence and the historical traceability sequence is calculated, constructing a local cost matrix. This matrix represents the matching cost between each node in the current sequence and each node in the historical sequence. Next, based on this local cost matrix, a dynamic programming algorithm is used to calculate the cumulative cost of all possible alignment paths from the start to the end of the sequence, and the path with the minimum cumulative cost is selected as the sequence normalization distance. Then, the reciprocal of the sum of this sequence normalization distance and a preset adjustment coefficient is calculated as the similarity between the two sequences. The higher the similarity, the better the current root cause localization sequence matches the historical traceability experience. When the calculated similarity is greater than a preset similarity threshold, such as 0.75, the newly added node connections in the current root cause localization sequence are extracted, and these newly added node connections are added as directed edges to the initial parameter relationship graph, thereby obtaining the updated target traceability graph.

[0058] In one possible implementation, the similarity between the root cause localization sequence and the historical tracing sequence in the historical production log data is calculated. When the similarity meets a preset similarity threshold, the node connection relationships in the root cause localization sequence are fused into the initial parameter relationship graph to obtain the target tracing graph. Specifically, this includes steps 1051-1053, as follows: Step 1051: Extract the feature parameters of each process node in the root cause localization sequence and the historical tracing sequence; calculate the spatial distance between the feature parameters of each process node in the root cause localization sequence and the historical tracing sequence, and construct the local cost matrix.

[0059] Specifically, the process begins by retrieving detailed information on each process node from the root cause localization sequence and the historical tracing sequence from the database. Then, for each process node, its feature parameters are extracted, including process parameters such as temperature, pH value, and soaking time; quality indicators such as bonding strength and copper layer thickness; and equipment operating status such as stirring speed and flow rate, among other multi-dimensional feature data. The feature parameters of each process node are represented as multi-dimensional feature vectors to facilitate spatial distance calculation. Next, the spatial distance between the feature parameter vectors of each process node in the root cause localization sequence and each process node in the historical tracing sequence is calculated. In this spatial distance calculation, the differences in each dimension of the two feature parameter vectors are calculated separately, the squares of these differences are summed, and the square root of the sum is taken. For example, if the feature parameter vector of the degumming process node in the root cause localization sequence is [temperature 60℃, pH 9.5, soaking time 5 minutes], and the feature parameter vector of the degumming process node in the historical tracing sequence is [temperature 58℃, pH 9.8, soaking time 4.5 minutes], then the square root of the sum of the squares of the differences in each dimension of the two vectors is calculated. Finally, the spatial distances between all the process node pairs are filled into a two-dimensional array according to the position of the nodes in the original sequence to construct a complete local cost matrix. The element value at position (i, j) in the matrix represents the spatial distance of the feature parameters between the i-th node in the root cause localization sequence and the j-th node in the historical tracing sequence.

[0060] Step 1052: Based on the local cost matrix, calculate the cumulative cost of all aligned paths from the start point to the end point of the sequence, and take the minimum cumulative cost as the sequence normalization distance.

[0061] Specifically, firstly, based on the constructed local cost matrix, a cumulative cost matrix of the same size is created to record the minimum cumulative cost from the start of the sequence to each position. The first element (0, 0) of the cumulative cost matrix is ​​initialized to the value of the corresponding position in the local cost matrix. Then, the remaining elements of the cumulative cost matrix are filled using a dynamic programming method. For the cumulative cost of position (i, j), three possible previous positions are considered: (i-1, j) represents skipping the i-th node in the root location sequence, (i, j-1) represents skipping the j-th node in the historical tracing sequence, and (i-1, j-1) represents aligning the i-th and j-th nodes of the two sequences. The cumulative cost from these three previous positions to (i, j) is calculated, and the minimum value is selected as the cumulative cost of position (i, j). For example, the cumulative cost at position (i, j) can be calculated by comparing the following three values: the cumulative cost at position (i-1, j) plus the skip cost, the cumulative cost at position (i, j-1) plus the skip cost, and the cumulative cost at position (i-1, j-1) plus the value at position (i, j) in the local cost matrix, and taking the minimum of these three. This process is repeated to fill all elements of the cumulative cost matrix until the lower right corner (m, n) is reached, where m is the length of the root location sequence and n is the length of the historical tracing sequence. Finally, the value of the lower right element of the cumulative cost matrix represents the minimum cumulative cost from the start to the end of the sequence, and this value is determined as the sequence normalization distance.

[0062] Step 1053: Calculate the reciprocal of the sum of the sequence regularization distance and the preset adjustment coefficient, and use the reciprocal as the similarity; when the similarity is greater than the preset similarity threshold, extract the newly added node connection relationship in the root location sequence, and add the newly added node connection relationship as a directed edge to the initial parameter relationship graph to obtain the target tracing graph.

[0063] Specifically, the process begins by obtaining the calculated sequence normalization distance value and simultaneously reading the pre-set adjustment coefficient value from the system. The adjustment coefficient is typically a positive number used to adjust the sensitivity and range of the similarity calculation. Then, the sequence normalization distance is added to the pre-set adjustment coefficient, resulting in a positive value greater than the sequence normalization distance. Next, the reciprocal of this positive value is calculated, i.e., 1 divided by (sequence normalization distance plus adjustment coefficient). This result is used as the similarity value between the root-source sequence and the historical traced sequence. For example, if the sequence normalization distance is 0.3 and the pre-set adjustment coefficient is 0.2, the similarity calculation is 1 divided by (0.3 + 0.2), which equals 1 divided by 0.5, resulting in 2.0. Subsequently, the calculated similarity value is compared with a pre-set similarity threshold in the system. If the similarity is greater than the pre-set threshold, subsequent map update steps continue; if it is not greater than the threshold, the current processing flow terminates. Once the similarity meets the threshold requirement, the process node connections in the root source localization sequence are analyzed and compared with existing node connections in the initial parameter relationship graph to identify new node connections present in the root source localization sequence but not in the initial graph. Each connection consists of a source node, a target node, and their associated attributes, represented as a directed edge. Finally, all identified new node connections are added as directed edges to the initial parameter relationship graph to form an updated target traceability graph, thereby enriching the representation of the association knowledge between parameters.

[0064] Step 106: Input the real-time monitoring data into the target traceability map, calculate the connectivity probability between each node in the target traceability map, generate a quality defect backtracking path based on the path corresponding to the highest connectivity probability, and extract upstream impact correlation indicators based on the quality defect backtracking path to generate a parameter adjustment instruction sequence.

[0065] In this embodiment, real-time monitoring data refers to the process parameters and quality characteristic data collected in real time at each stage of the current production process, used to reflect the immediate status of the production line. Connectivity probability represents the probability value of real-time monitoring data propagating along a specific path in the target traceability map, used to quantify the likelihood of a specific defect propagation path. The quality defect backtracking path refers to the parameter propagation link most likely to lead to the current quality defect, determined based on connectivity probability calculations, used to locate the root cause of the defect. The upstream impact correlation index represents the deviation of upstream process parameters in the quality defect backtracking path and its degree of impact on downstream quality, used to guide parameter adjustments. The parameter adjustment instruction sequence refers to the process parameter adjustment scheme generated based on the upstream impact correlation index, used to guide production equipment to perform automatic or manual parameter adjustments.

[0066] Specifically, firstly, real-time monitoring data containing the deviation of process parameters for the current batch is mapped to the corresponding nodes in the target traceability graph, and the current copper plating defects, such as insufficient bonding force or uneven coverage, are located, along with the corresponding target defect nodes. Then, based on the directed edge connections in the target traceability graph, all connected branches pointing from each upstream process node to the target defect node are traversed, extracting possible connected node sequences as candidate backtracking paths. Next, combining the historical confidence of each directed edge in the target traceability graph, the conditional probability of real-time monitoring data transmission between adjacent nodes in each candidate backtracking path is calculated. The conditional probability calculation is based on Bayesian principles, representing the probability that a downstream node will exhibit a specific state given the parameter state of an upstream node. Subsequently, the conditional probabilities between all adjacent nodes on each candidate backtracking path are multiplied to obtain the overall connectivity probability of that candidate path. Then, the connectivity probability values ​​of all candidate backtracking paths are compared, and the path with the highest connectivity probability is selected as the final quality defect backtracking path. Finally, the parameter deviation of the main root cause process and the constraint margin of the current copper plating process are extracted from this backtracking path as upstream influence correlation indicators. If the copper plating constraint margin is lower than the preset acceptable lower limit, the influence gradient of the parameter deviation on the copper plating constraint margin is further calculated, i.e., the change in margin caused by a unit parameter change, and the required parameter adjustment range is calculated based on this influence gradient. The influence gradient is determined through regression analysis or local sensitivity analysis of historical data. Specifically, paired samples of the main root cause process parameters and copper plating constraint margins are extracted from historical production data to establish a parameter-margin regression model, such as linear regression or locally weighted regression, and the slope coefficient of the model is the influence gradient; or several nearby historical samples are selected near the current operating point, the ratio of the change in copper plating constraint margin to the change in the main root cause process parameters is calculated, and the average value is taken as the local influence gradient. For cases of multi-parameter coupled influence, partial derivative calculation is used to obtain the individual influence gradient of the target parameter on the margin while keeping other parameters constant, or the partial regression coefficients of each parameter are obtained through a multivariate regression model as their respective influence gradient values. The tuning of the influence gradient adopts a dynamic update mechanism, and the initial gradient value is obtained by offline analysis of the parameter-margin relationship of historical batch data. During production, whenever parameters are adjusted, the actual changes in the main source process parameters before and after the adjustment, along with the corresponding changes in copper plating constraint margins, are recorded. The measured gradient value is calculated and compared with the model-predicted gradient value. If the deviation exceeds the set range, the gradient value is corrected online using an exponentially weighted moving average or recursive least squares method. The new gradient value is obtained by fusing historical gradient values ​​and measured gradient values ​​with a certain weighting coefficient. At the same time, a confidence interval for the gradient value is established. When the cumulative sample size reaches the statistical significance requirement, the nominal value and fluctuation range of the gradient are updated to ensure that the gradient calculation can adapt to the influence of factors such as process drift, equipment aging, and changes in raw material batches, maintaining the accuracy and stability of parameter adjustments.The main root cause process is identified as the target adjustment node, and the identification information of this node is combined with the calculated adjustment range to generate a parameter adjustment instruction sequence, which is used to guide the production equipment to optimize parameters.

[0067] In one possible implementation, real-time monitoring data is input into the target traceability map, the connectivity probability between nodes in the target traceability map is calculated, and a quality defect backtracking path is generated based on the path corresponding to the highest connectivity probability. Specifically, this includes steps 1061-1063, as follows: Step 1061: Map the real-time monitoring data containing the process deviation of the current batch to the corresponding node of the target traceability graph, and locate the target defect node corresponding to the current copper plating defect; according to the directed edge connection relationship in the target traceability graph, traverse all connected branches from each upstream process node to the target defect node, and use the extracted connected node sequence as a candidate backtracking path.

[0068] Specifically, the process parameter data for the current batch collected by the real-time monitoring system is first categorized and organized according to process type and parameter attributes, and then matched with nodes in the target traceability graph. During matching, the real-time monitoring data is accurately mapped to the corresponding node positions in the graph based on information such as parameter name, process type, and timestamp. Then, by analyzing the defect characteristics of the current copper plating layer, such as poor copper plating and insufficient adhesion, the target defect nodes matching these defect characteristics are located in the target traceability graph. Next, starting from the target defect node, the graph is traversed backwards along the directed edges. During the traversal, each incoming edge pointing to the target defect node is checked, and all incoming edges from the starting nodes of these incoming edges are recursively checked until the source node with no incoming edges is reached. Each time a complete path from the source node to the target defect node is encountered, all nodes on that path are organized into a sequence according to the directed edge connection order, and saved as a candidate backtracking path. For example, if there is a connected path from the descaling node through the micro-etching node to the copper plating defect node, then "descaling to micro-etching to copper plating defect" is considered a candidate backtracking path. In this way, a set of all candidate backtracking paths that may lead to the target defect is finally obtained, providing a basis for subsequent probability calculations and path selection.

[0069] Step 1062: Combine the historical confidence of each directed edge in the optimized tracing graph to calculate the conditional probability of real-time monitoring data being transmitted between adjacent nodes of the candidate backtracking path.

[0070] Specifically, firstly, historical confidence information for each directed edge is extracted from the optimized traceability graph. This confidence level is derived from statistical analysis of historical production data and reflects the intensity of the impact of upstream node parameter changes on downstream node parameters. Then, each path in the candidate backtracking paths is analyzed segment by segment. For each pair of adjacent nodes on a path, the historical confidence level of the directed edge connecting them is obtained as a prior probability. Next, combined with the real-time monitoring data of the current batch, the conditional probability between these adjacent nodes is calculated. During the calculation, the degree of parameter deviation of the upstream node is first determined and compared with the impact on downstream nodes under similar deviation conditions in historical data. For continuous parameters, the probability density of the downstream node parameter taking its current measured value under the given upstream node parameter value is calculated. For discrete parameters, the frequency probability of the downstream node exhibiting its current state under the current state of the upstream node is calculated. Taking into account both the prior probability and the consistency with the current data, the conditional probability between adjacent nodes is obtained. For example, if the current pH of the micro-etching solution is 8.2, historical data shows that the probability of a decrease in the adsorption rate of the activation solution when the pH is in the range of 8.0-8.3 is 0.65, and the historical confidence of this directed edge in the optimized tracing graph is 0.8, then the calculated conditional probability is approximately 0.52. Using this method, the conditional probability between each pair of adjacent nodes on the candidate backtracking path is calculated to prepare for the calculation of the subsequent path connectivity probability.

[0071] Step 1063: Multiply the conditional probabilities between all adjacent nodes on each candidate backtracking path to obtain the connectivity probability of the candidate backtracking path; compare the connectivity probabilities of each candidate backtracking path and select the candidate backtracking path with the highest connectivity probability as the quality defect backtracking path.

[0072] Specifically, the conditional probabilities of each candidate backtracking path are first processed. For each candidate backtracking path, starting from the source node, the conditional probability values ​​between all adjacent node pairs on the path are sequentially extracted. Then, the conditional probability values ​​of all adjacent node pairs on this path are multiplied together, that is, the conditional probability between the first pair of nodes is multiplied by the conditional probability between the second pair of nodes, then by the conditional probability between the third pair of nodes, and so on, until the conditional probability between the last pair of adjacent nodes on the path is multiplied. This multiplication operation is based on the chain rule of parameter propagation, which can calculate the comprehensive probability of continuous propagation of parameter deviations along the entire path. For example, if a candidate backtracking path contains three pairs of adjacent nodes with conditional probabilities of 0.8, 0.7, and 0.6, the connectivity probability of this path is 0.336. After calculating the connectivity probabilities of all candidate backtracking paths, these connectivity probability values ​​are sorted in descending order, and the connectivity probability values ​​of each path are compared. Finally, the candidate backtracking path with the highest connectivity probability is selected as the final quality defect backtracking path. For example, if there are three candidate backtracking paths with connectivity probabilities of 0.42, 0.58 and 0.37 respectively, then the path with connectivity probability of 0.58 is selected as the quality defect backtracking path, as this path is most likely to explain the cause of the current copper plating defect.

[0073] In one possible implementation, a parameter adjustment instruction sequence is generated by extracting upstream impact correlation indicators based on the quality defect backtracking path, specifically including steps 1064-1066, as follows: Step 1064: Extract the parameter deviation and copper plating constraint margin of the main root cause process from the quality defect backtracking path as upstream influence correlation indicators.

[0074] Specifically, the process begins by analyzing parameter anomalies at each process node within the identified quality defect backtracking path. Based on the node's position, influence weight, and degree of anomaly within the backtracking path, the process node with the most significant impact on the current copper plating defect is identified and designated as the primary root cause process. The determination of the primary root cause process can be based on graph theory characteristics such as the node's in-degree, out-degree, and critical position within the path, or it can be comprehensively evaluated in conjunction with the severity of the node's parameter anomalies. Next, the actual values ​​of key process parameters are extracted from the real-time monitoring data of the primary root cause process and compared with the standard value or historical best value of the parameter to calculate the deviation. For example, if the standard value of the micro-etching solution concentration is 120 g / L and the actual value is 105 g / L, the parameter deviation is -15 g / L. Simultaneously, by analyzing the differences between the measured values ​​of various quality indicators of the copper plating layer and the corresponding lower limits of the acceptable standards, the copper plating constraint margin for the current batch of products is calculated. The identified primary root cause process parameter deviations and the calculated copper plating constraint margins are used as upstream influence correlation indicators for subsequent parameter adjustment analysis.

[0075] Step 1065: Determine whether the copper plating constraint margin is lower than the preset acceptable lower limit; if the copper plating constraint margin is lower than the preset acceptable lower limit, calculate the gradient of the influence of parameter deviation on the copper plating constraint margin.

[0076] Specifically, the first step is to obtain the calculated copper plating constraint margin value and compare it with a pre-set acceptable lower limit. The pre-set acceptable lower limit is the minimum acceptable margin value determined based on product quality standards and process stability requirements, typically set as a safe range value above the standard acceptance line. When the current copper plating constraint margin is found to be lower than the pre-set acceptable lower limit, the parameter adjustment calculation process is triggered. Then, based on historical production data and the current process status, a mapping relationship model between the main root cause process parameters and the copper plating constraint margin is established. By analyzing the changes in the copper plating constraint margin under different parameter values, numerical differentiation or regression analysis methods are applied to calculate the gradient of the influence of parameter deviation on the copper plating constraint margin. During the calculation, the main root cause process parameters are used as independent variables, and the copper plating constraint margin is used as the dependent variable. Several historical data points are taken near the current operating point, and the gradient value is obtained using the difference quotient method or derivative fitting method. For multi-parameter cases, a gradient vector of a multivariate function can be constructed, where each component represents the partial derivative value of each parameter with respect to the margin.

[0077] Step 1066: Calculate the adjustment range of the compensation parameter deviation based on the influence gradient, determine the main root cause process as the target adjustment node, combine the identification information of the target adjustment node with the adjustment range, and generate a parameter adjustment instruction sequence.

[0078] Specifically, firstly, based on the calculated influence gradient, the required adjustment range for compensating parameter deviations is determined. The adjustment range calculation considers the difference between the current copper plating constraint margin and the target margin, as well as the sensitivity of the parameter deviation to the margin. The optimal adjustment range is obtained by solving the functional relationship between the parameter adjustment amount and the margin increase amount. The adjustment range calculation must consider the effective range of parameter adjustment and process constraints to ensure that the adjusted parameter values ​​remain within the process's allowable range. Then, the identified main root cause process is marked as the target adjustment node, and its identification information, such as process number, equipment ID, and parameter name, is extracted. Finally, the identification information of the target adjustment node and the calculated adjustment range are structurally combined to generate a standard-format parameter adjustment instruction sequence. Each instruction in the sequence clearly specifies the adjustment object, adjustment parameter, adjustment direction, and adjustment amount, forming a complete process parameter adjustment scheme. For complex adjustment scenarios, multiple adjustment instructions may be generated and ordered by priority to form an ordered adjustment instruction sequence.

[0079] It should be noted that the models involved in this application mainly include joint probability distribution models, Mahalanobis distance anomaly detection models, bias propagation gain models, and connectivity probability inference models. Model tuning employs a combination of offline training and online updates. During the offline training phase, historical production data is used to establish the initial parameters of each model, such as covariance matrix, marginal probability distribution, and directed edge confidence. During the online operation phase, once a certain amount of new production batch data has accumulated, the model parameters are periodically recalculated or the model is updated using a sliding window approach. Simultaneously, a mechanism for monitoring the rate of change of model parameters is set up. When the parameter changes exceed a set range, model retraining is triggered to ensure that the model can adapt to the influence of factors such as changes in process conditions and equipment aging, maintaining traceability accuracy.

[0080] In the above embodiments, basic copper plating process parameter optimization functions were achieved by generating quality defect backtracking paths and constructing parameter adjustment instruction sequences. To further improve the accuracy and reliability of abnormal parameter identification, a multi-dimensional correlation analysis model of process parameters, quality characteristics, and historical operating conditions was established. This application also provides a method for tracing the entire horizontal copper plating production process. This method analyzes the statistical distance characteristics between the parameter influence intensity and historical operating condition samples, constructs an anomaly point detection and parameter correlation mapping mechanism in a multi-dimensional feature space, and performs anomaly sample identification and anomaly parameter extraction based on Mahalanobis distance. This enables the system to more accurately identify key process parameters that cause abnormal hole wall roughness in complex environments with multi-parameter interactive influences. The following is a combination of... Figure 2 The present application describes a method for tracing the entire production process of horizontal copper plating: Please see Figure 2 This is another schematic diagram of a method for tracing the entire process of horizontal copper plating production, as described in this application.

[0081] Step 201: If the parameter influence intensity index exceeds the preset threshold, then extract the hole wall roughness feature from the feature vector.

[0082] Specifically, the process first acquires the measured parameter influence intensity index under the current operating conditions. This index characterizes the degree to which changes in process parameters affect the borehole wall quality. Then, this index is compared with a pre-set threshold, which is determined based on historical data statistics and process requirements. When the value of the parameter influence intensity index exceeds the pre-set threshold, the feature extraction process is triggered. Next, the collected feature vector data is accessed. The feature vector contains various process quality features, such as borehole diameter, hole position accuracy, and borehole wall roughness. Based on the structural definition of the feature vector, the index position of the borehole wall roughness feature in the vector is determined, and the corresponding subset of borehole wall roughness features is extracted from the feature vector. The extraction process may involve data format conversion and dimension reorganization to ensure that the extracted borehole wall roughness features meet the requirements of subsequent processing.

[0083] Step 202: Obtain a reference domain containing multiple historical working condition samples; calculate the Mahalanobis distance between the hole wall roughness features and each historical working condition sample in the reference domain to form a distance matrix.

[0084] Specifically, firstly, a reference domain containing multiple historical operating condition samples is obtained. This reference domain consists of operating condition data collected under normal production conditions in the past. The construction of the reference domain can be based on screening criteria such as time windows, quality grades, or process conditions to ensure the representativeness and comparability of the reference samples. Then, Mahalanobis distance is calculated between the extracted hole wall roughness features and each historical operating condition sample in the reference domain. The calculation process first estimates the covariance matrix of the reference domain data to capture the correlation structure between features of each dimension. Next, a standardized Mahalanobis distance value is calculated between the current hole wall roughness feature and each historical sample in the reference domain. This calculation specifically includes: based on the obtained reference domain covariance matrix and its inverse matrix, for each historical operating condition sample in the reference domain, the difference vector between its feature vector and the current hole wall roughness feature vector to be detected is calculated; then, using the inverse of the covariance matrix as the weight matrix, the difference vector is transposed and sequentially multiplied with the covariance inverse matrix and the original difference vector to obtain a matrix multiplication operation reflecting both. In a multidimensional feature space, a scalar value of the distance is statistically calculated. Taking the square root of this scalar value yields a Mahalanobis distance value. This calculation is repeated for all historical samples in the reference domain to obtain the Mahalanobis distance between the current feature and each historical sample. Finally, these distance values ​​are arranged in the order of the historical sample numbers to form a distance matrix in the form of a one-dimensional array or column vector. Each element represents the standardized distance between the current hole wall roughness feature and the corresponding historical sample. The standardization is achieved by automatically eliminating dimensional differences between feature dimensions and considering the correlation between dimensions using the inverse of the covariance matrix as weights, making the distances between different samples statistically comparable. For a reference domain containing n samples, n Mahalanobis distance values ​​need to be calculated. These n distance values ​​are arranged according to the correspondence between the current feature and each historical sample to form a distance matrix. The distance matrix can be represented as an n-dimensional vector or an n×1 matrix, where each element represents the Mahalanobis distance between the current hole wall roughness feature and the corresponding historical sample.

[0085] In one possible implementation, the Mahalanobis distance between the hole wall roughness features and each historical working condition sample in the reference domain is calculated to form a distance matrix, specifically including steps 2021-2023, as follows: Step 2021: Calculate the covariance matrix of the historical operating condition samples in the reference domain, and solve for the inverse of the covariance matrix.

[0086] Specifically, firstly, borehole wall roughness feature data for all historical working condition samples in the reference domain are collected. These data form a multi-row, multi-column matrix, where each row represents a historical sample and each column represents a feature dimension. Next, the covariance values ​​between each feature dimension in this data matrix are calculated. During the calculation, for each pair of feature dimensions, their respective means are calculated. Then, for each sample, the product of the deviations of each feature from its mean is calculated. Finally, these products for all samples are summed and divided by the sample size minus one to obtain the covariance of this pair of feature dimensions. This calculation process is repeated for all feature dimension pairs, and the obtained covariance values ​​are filled into the covariance matrix according to their corresponding positions. The covariance matrix is ​​a square matrix whose dimensions are equal to the number of feature dimensions and it is symmetric. After calculating the covariance matrix, the inverse matrix is ​​obtained through matrix inversion. Matrix inversion can be achieved using numerical algorithms such as Gaussian-Jordan elimination and LU decomposition. If the covariance matrix is ​​close to singular and its determinant is close to zero, regularization techniques may be needed to ensure the numerical stability of the inverse matrix. The resulting inverse matrix will be used as weights for Mahalanobis distance calculation in subsequent steps.

[0087] Step 2022: Using the inverse matrix as weights, calculate the quadratic difference between the hole wall roughness characteristics and each historical working condition sample.

[0088] Specifically, the current hole wall roughness feature to be detected is first represented as a feature vector, which contains roughness index data in multiple dimensions. Then, the corresponding feature vector for each historical working condition sample is extracted one by one from the reference domain. For each historical sample, the difference between its feature vector and the currently detected feature vector is calculated to obtain the difference vector. Next, the inverse of the calculated covariance matrix is ​​used to perform mathematical operations with the difference vector. The specific operation process is as follows: first, the difference vector is transposed; then, the transposed difference vector is multiplied on the left by the inverse of the covariance matrix; finally, the result is multiplied on the right by the original difference vector. The result of this series of matrix multiplication operations is the quadratic difference value, which is a scalar. The above calculation process is performed for each historical working condition sample in the reference domain to obtain the quadratic difference value between the current hole wall roughness feature and each historical working condition sample. These difference values ​​will be used as Mahalanobis distance in the next step.

[0089] Step 2023: Fill the corresponding element positions in the matrix with the quadratic difference value as the Mahalanobis distance to generate the distance matrix.

[0090] Specifically, the dimension of the distance matrix is ​​first determined based on the number of historical operating condition samples in the reference domain. If the reference domain contains n historical operating condition samples, the distance matrix will be an n×1 column vector or a 1×n row vector, depending on the implementation. Then, the square root of each quadratic difference value is used as the Mahalanobis distance. For each historical operating condition sample in the reference domain, its corresponding Mahalanobis distance is filled into the corresponding element of the distance matrix according to the sample index order. During the filling process, it is necessary to ensure that the correspondence between the Mahalanobis distance and the historical samples is accurate, ensuring that each element in the distance matrix correctly reflects the statistical distance between the current hole wall roughness characteristics and the specific historical operating condition sample. After filling all Mahalanobis distances, the distance matrix is ​​formatted as necessary to ensure that its data structure meets the requirements of subsequent processing. The final generated distance matrix will serve as the data source for extracting potential outliers, providing a basis for identifying anomalous parameter sets.

[0091] Step 203: Extract sample points whose distance values ​​are greater than the preset distance judgment threshold from the distance matrix as potential anomalies, and summarize the parameters corresponding to the potential anomalies into an anomaly parameter set.

[0092] Specifically, firstly, distance values ​​are extracted from the generated distance matrix. Each distance value represents the degree of statistical difference between the current hole wall roughness characteristics and a historical working condition sample. Then, these distance values ​​are compared with a pre-determined preset distance judgment threshold. The preset distance judgment threshold can be determined through statistical methods such as confidence intervals based on normal distribution, historical data analysis, or expert experience. For each distance value, if it exceeds the preset distance judgment threshold, the corresponding sample point is marked as a potential anomaly. Next, all samples marked as potential anomalies are collected, and the corresponding process parameter information is extracted, including parameter name, parameter value, and the degree of deviation from the normal range. Finally, all relevant parameters extracted from the potential anomalies are summarized and categorized. Possible categorization methods include grouping by parameter type, sorting by influence degree, or dividing by process stage, ultimately forming a structured set of abnormal parameters.

[0093] It should be noted that the preset threshold is determined by statistically analyzing the relative entropy distribution of drilling parameters and degumming effect in historical production data. Specifically, sample data from several recent batches of normal and abnormal production are collected, and the parameter influence intensity index values ​​are calculated for each batch. Frequency distribution histograms are plotted, and the boundary point that can effectively distinguish between normal and abnormal operating conditions is selected as the threshold. Typically, the 95th or 98th percentile of the relative entropy distribution under normal operating conditions is chosen to ensure that significant parameter interaction effects are captured while avoiding excessive false alarms.

[0094] The preset similarity threshold is set based on the sequence similarity distribution of historically verified defect cases. Confirmed root cause quality defect cases are extracted from the historical production log database, and the similarity value between their root cause localization sequence and the historical traceability sequence is calculated. The distribution range of these similarities is statistically analyzed, and the minimum similarity value that covers most successful traceability cases is selected as the threshold. This threshold is generally set between the 10th and 20th percentiles of the historical successful case similarity distribution, balancing the accuracy and adaptability of the map update.

[0095] The preset numerical intervals are divided based on the actual value range and statistical distribution characteristics of process parameters and quality features. For the drilling parameter component and the degumming effect component, the maximum, minimum, and standard deviation of each parameter in historical data are first obtained. Then, the parameter value range is divided into several intervals using the equal-width division method or the equal-frequency division method. The number of intervals is usually set to 5 to 10 to ensure that there is a sufficient sample size in each interval to support probability statistics. At the same time, the interval division needs to cover the fluctuation range allowed by the process specifications, and the intervals should be appropriately densified near the key control points to improve resolution.

[0096] The preset distance judgment threshold is determined by the statistical distribution characteristics of Mahalanobis distance. Under the assumption that the samples in the reference domain follow a multivariate normal distribution, the square of the Mahalanobis distance follows a chi-square distribution. The critical value is obtained by consulting the chi-square distribution table based on the number of feature dimensions and the confidence level. Usually, the chi-square critical value corresponding to the 95% or 99% confidence level is selected as the judgment threshold. Alternatively, the mean of the Mahalanobis distance between samples in the reference domain is calculated plus 2 or 3 times the standard deviation as the threshold. Sample points exceeding this threshold are judged as outliers that deviate from the standard operating conditions.

[0097] The preset adjustment coefficient is used to adjust the sensitivity and numerical range when converting sequence regularization distance to similarity. The setting of the adjustment coefficient needs to consider the typical range of sequence regularization distance and the desired similarity range. Usually, by analyzing the distribution of regularization distance of historical sequence pairs, a coefficient value that can reasonably distribute the similarity between 0 and 1 or between 0 and 2 is selected. It is generally set to 0.5 to 1 times the mean of the regularization distance of historical sequences, or the coefficient value that optimizes the similarity discrimination effect can be selected through cross-validation.

[0098] The preset acceptable lower limit is set based on product quality standards and process stability requirements. First, determine the quality acceptable line for the copper plating constraint margin, which is the minimum margin value that meets product specifications. Then, add a safety margin based on the acceptable line. This safety margin comprehensively considers the process fluctuation range, measurement error range, and quality risk level, and is usually set to 1.1 to 1.3 times the quality acceptable line value, or add 1 to 2 times the standard deviation of the margin fluctuation under historical normal operating conditions to the acceptable line, ensuring sufficient warning time and adjustment space before parameter adjustment.

[0099] The following describes a horizontal copper plating production process traceability system from the perspective of hardware processing, according to an embodiment of this invention. Please refer to [link / reference]. Figure 3 This is a schematic diagram of the structure of a horizontal copper plating production process traceability system in an embodiment of this application.

[0100] It should be noted that, Figure 3 The structure of the horizontal copper plating production process traceability system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0101] like Figure 3 As shown, a traceability system for the entire horizontal copper plating production process includes a central processing unit (CPU) 301. This CPU can perform various appropriate actions and processes based on programs stored in a read-only memory (ROM) 302 or loaded from storage section 308 into a random access memory (RAM) 303, such as the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. Input / output (I / O) interfaces 305 are also connected to the bus 304.

[0102] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including liquid crystal displays (LCDs), audio output devices, indicator lights, etc.; storage section 308 including hard disks, etc.; and communication section 309 including network interface cards such as LANs, Local Area Networks, cards, modems, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

Claims

1. A method for tracing the entire production process of horizontal copper plating, characterized in that, The method includes: Obtain the feature vectors corresponding to cross-process parameter correlation data; Based on the feature vector, borehole parameter components and desizing effect components are extracted, the joint probability distribution between the borehole parameter components and desizing effect components is calculated, and the parameter influence intensity index is determined based on the joint probability distribution. If the influence of the parameter on the intensity index exceeds a preset threshold, then the distance matrix between the hole wall roughness feature in the feature vector and the historical working condition sample is calculated, and potential abnormal points deviating from the standard working condition are screened out based on the distance matrix to obtain an abnormal parameter set. Extract the process nodes corresponding to the copper plating defects from the abnormal parameter set, calculate the deviation propagation gain between adjacent process nodes according to the process sequence, construct the influence chain based on the deviation propagation gain, and determine the root cause localization sequence. Calculate the similarity between the root cause location sequence and the historical traceability sequence in the historical production log data. When the similarity meets a preset similarity threshold, merge the node connection relationship in the root cause location sequence into the initial parameter relationship graph to obtain the target traceability graph. The real-time monitoring data is input into the target traceability map, the connectivity probability between each node in the target traceability map is calculated, a quality defect backtracking path is generated based on the path corresponding to the highest connectivity probability, and an upstream impact correlation index is extracted based on the quality defect backtracking path to generate a parameter adjustment instruction sequence.

2. The method according to claim 1, characterized in that, The step of obtaining the feature vector corresponding to the cross-process parameter correlation data includes: Obtain cross-process parameter correlation data for each process in the pre-production stage, wherein the cross-process parameter correlation data includes measured values ​​of process parameters and measured values ​​of quality characteristics; Based on the process specification settings of each process, the normalized deviation of the measured values ​​of the process parameters is calculated; The normalized deviation is concatenated with the corresponding measured quality feature values ​​to generate the feature vector.

3. The method according to claim 1, characterized in that, The step of extracting borehole parameter components and desizing effect components based on the feature vector, calculating the joint probability distribution between the borehole parameter components and the desizing effect components, and determining the parameter influence intensity index based on the joint probability distribution includes: The drilling parameter components characterizing the drilling speed and feed rate, and the adhesive removal effect component characterizing the residual adhesive thickness are separated from the feature vector. The co-occurrence frequency of the drilling parameter components and the degumming effect components within a preset value range is statistically analyzed, and the co-occurrence frequency is converted into a joint probability distribution. The independent occurrence frequencies of the drilling parameter components and the degumming effect components in their respective component value ranges are statistically analyzed, and the independent occurrence frequencies are converted into corresponding marginal probability distributions. The marginal probability distributions of the drilling parameter components and the marginal probability distributions of the descaling effect components are multiplied to generate independent probability distributions. Calculate the relative entropy between the joint probability distribution and the independent probability distribution, and use the relative entropy as an index of the influence intensity of the parameter.

4. The method according to claim 1, characterized in that, If the influence of the parameter on the intensity index exceeds a preset threshold, then the distance matrix between the hole wall roughness feature in the feature vector and the historical working condition sample is calculated. Based on the distance matrix, potential anomalies deviating from the standard working condition are screened out to obtain an abnormal parameter set, including: If the influence of the parameter on the strength index exceeds a preset threshold, then the hole wall roughness feature in the feature vector is extracted; Obtain a reference domain containing multiple historical operating condition samples; Calculate the Mahalanobis distance between the hole wall roughness feature and each of the historical working condition samples in the reference domain to form the distance matrix; Sample points whose distance values ​​are greater than a preset distance judgment threshold are extracted from the distance matrix as potential anomalies, and the parameters corresponding to the potential anomalies are summarized into the anomaly parameter set.

5. The method according to claim 4, characterized in that, The calculation of the Mahalanobis distance between the hole wall roughness feature and each of the historical working condition samples in the reference domain, forming the distance matrix, includes: Calculate the covariance matrix of the historical working condition samples in the reference domain, and solve for the inverse of the covariance matrix; Using the inverse matrix as weights, calculate the quadratic difference value between the hole wall roughness feature and each of the historical working condition samples; The quadratic difference value is used as the Mahalanobis distance and filled into the corresponding element position of the matrix to generate the distance matrix.

6. The method according to claim 1, characterized in that, The process of extracting the process nodes corresponding to copper plating defects from the abnormal parameter set, calculating the deviation propagation gain between adjacent process nodes according to the process sequence, constructing an influence chain based on the deviation propagation gain, and determining the root cause localization sequence includes: Traverse the set of abnormal parameters, and according to the preset defect parameter mapping table, filter out the target parameters that match the copper bonding strength defect from the set of abnormal parameters, and take the process where the target parameter is located as the process node. Obtain the parameter changes and quality characteristic changes between adjacent process nodes; The ratio of the change in the quality characteristic to the change in the parameter is used as the deviation transfer gain. The corresponding process nodes are connected in series according to the order of increasing deviation propagation gain to generate the influence chain, and the node sequence in the influence chain is used as the root cause localization sequence.

7. The method according to claim 1, characterized in that, The calculation of the similarity between the root cause localization sequence and the historical tracing sequence in the historical production log data, and when the similarity meets a preset similarity threshold, the node connection relationships in the root cause localization sequence are fused into the initial parameter relationship graph to obtain the target tracing graph, includes: Extract the feature parameters of each process node in the root cause localization sequence and the historical tracing sequence; Calculate the spatial distance between the feature parameters of each process node in the root cause localization sequence and the historical tracing sequence, and construct a local cost matrix; Based on the local cost matrix, calculate the cumulative cost of all aligned paths from the start point to the end point of the sequence, and take the minimum cumulative cost as the sequence normalization distance; Calculate the reciprocal of the sum of the sequence normalization distance and the preset adjustment coefficient, and use the reciprocal as the similarity. When the similarity is greater than a preset similarity threshold, the newly added node connection relationship in the root location sequence is extracted, and the newly added node connection relationship is added as a directed edge to the initial parameter relationship graph to obtain the target tracing graph.

8. The method according to claim 1, characterized in that, The step of inputting real-time monitoring data into the target traceability map, calculating the connectivity probability between nodes in the target traceability map, and generating a quality defect backtracking path based on the path corresponding to the highest connectivity probability includes: The real-time monitoring data containing the process deviation of the current batch is mapped to the corresponding node of the target traceability map, and the target defect node corresponding to the current copper plating defect is located. Based on the directed edge connection relationship in the target traceability graph, traverse all connected branches from each upstream process node to the target defect node, and use the extracted connected node sequence as a candidate backtracking path; Based on the historical confidence of each directed edge in the target tracing graph, the conditional probability of the real-time monitoring data being transmitted between adjacent nodes of the candidate backtracking path is calculated. The connectivity probability of the candidate backtracking path is obtained by multiplying the conditional probabilities between all adjacent nodes on each candidate backtracking path. By comparing the connectivity probabilities of each candidate backtracking path, the candidate backtracking path with the highest connectivity probability is selected as the quality defect backtracking path.

9. The method according to claim 1, characterized in that, The step of extracting upstream impact correlation indicators based on the quality defect backtracking path to generate a parameter adjustment instruction sequence includes: The parameter deviation and copper plating constraint margin of the main root cause process are extracted from the quality defect backtracking path as the upstream influence correlation index. Determine whether the copper plating constraint margin is lower than the preset acceptable lower limit; If the copper plating constraint margin is lower than the preset acceptable lower limit, then calculate the gradient of the influence of the parameter deviation on the copper plating constraint margin. The adjustment range for compensating the parameter deviation is calculated based on the influence gradient. The main root cause process is determined as the target adjustment node. The identification information of the target adjustment node is combined with the adjustment range to generate the parameter adjustment instruction sequence.

10. A traceability system for the entire process of horizontal copper plating production, characterized in that, The horizontal copper plating production process traceability system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the horizontal copper plating production process traceability system to perform the method as described in any one of claims 1-9.

Images