Big data-based circuit board production data management method and system

Abstract

The invention relates to the technical field of electronic data processing, and discloses a circuit board production data management method and system based on big data, and the method comprises the steps: obtaining the original production data of a circuit board, and constructing each data point into a multi-dimensional feature vector according to a timestamp, a working procedure, and a data physical attribute; clustering the multi-dimensional feature vectors into a plurality of process feature data layers by adopting a clustering algorithm and utilizing a distance metric function; calculating a cross-layer transfer entropy, and when the cross-layer transfer entropy is greater than a preset threshold and the direction is consistent with the process flow, establishing a directed dependency edge, and constructing a non-cyclic directed graph; and extracting production data of the circuit board to be monitored, mapping the production data to the process characteristic data layer, calculating mutual information among real-time data, calculating a judgment threshold value, and judging that data association is abnormal when the mutual information is lower than the judgment threshold value. According to the method, structured layering can be carried out on production data, causal dependence between layers is quantified, and early warning of potential quality problems is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic data processing, in particular to a circuit board production data management method and system based on big data. BACKGROUND

[0002] The manufacturing process of printed circuit board is extremely complex, covering dozens of precise physical and chemical processes such as drilling, electroplating, etching and lamination. With the development of industrial automation and intelligentization, the manufacturing execution system (MES), equipment sensors and automatic optical inspection (AOI) systems on the production line will generate massive, multi-dimensional data in real time, which records key information such as equipment status, process parameters, material characteristics and product quality.

[0003] Traditional data management adopts linear storage method based on time series or production batch, simply stacking data of different processes and different physical properties, which is difficult to identify the internal structure and correlation of data in time and process flow dimensions. When analyzing, statistical process control and other methods are usually used to monitor individual parameters of a single process independently. Small fluctuations in upstream processes may affect downstream processes through complex transmission effects, leading to product defects.

[0004] When detecting anomalies, if the parameters of two processes are within the specification range, but the historical strong correlation between them suddenly weakens, there may be potential process anomalies. In addition, although simple correlation analysis can find the relationship between variables, it cannot judge the directionality and transmission strength of the impact, so it is difficult to build a data model that can reflect the causal transmission chain in the process flow.

[0005] Therefore, it is necessary to design a production data management method that can automatically structure and layer production data, quantify causal dependence between levels, and implement anomaly detection based on this, in order to improve the intelligent control level of PCB production process. SUMMARY

[0006] The present application provides a circuit board production data management method and system based on big data to solve the problem of being unable to structure and layer production data, quantify causal dependence between levels in the prior art.

[0007] In a first aspect, the present application provides a circuit board production data management method based on big data, comprising the following steps: Obtain the original production data in the circuit board production process, and construct each data point into a multi-dimensional feature vector according to the time stamp, the process to which it belongs, and the data physical property; Use a clustering algorithm based on density peaks, and use a distance measurement function that represents time proximity and upstream and downstream relationships in the process as nonlinear weights to cluster the multi-dimensional feature vector into multiple process feature data layers; calculating cross-layer transfer entropy between any two process feature data layers, when the cross-layer transfer entropy is greater than a preset threshold and the direction is consistent with the process flow, establishing a directed dependency relationship edge between the two process feature data layers, and constructing a non-cyclic directed graph representing the intrinsic causal association of data by taking the cross-layer transfer entropy as the edge weight; extracting production data of a to-be-monitored circuit board production batch and mapping to the corresponding process feature data layer, for any pair of process feature data layers connected by a directed dependency relationship edge in the non-cyclic directed graph, calculating the mutual information between real-time data in the process feature data layer, and calculating a decision threshold positively correlated with the historical edge weight of the directed dependency relationship edge, when the mutual information is lower than the decision threshold, determining that the data association of the to-be-monitored circuit board production batch is abnormal.

[0008] Preferably, the data points are constructed into a multi-dimensional feature vector according to the timestamp, the process to which the data point belongs, and the data physical property, including: converting the timestamp into cumulative seconds from the start of the production batch; converting the process to which the data point belongs into a binary vector using one-hot encoding to obtain a process vector; obtaining standardized physical property values after Z-score standardization processing of the data physical property, the data physical property including temperature, pressure, and chemical concentration; and concatenating the cumulative seconds, the process vector, and the standardized physical property values in order to obtain the multi-dimensional feature vector.

[0009] Preferably, the time proximity and the upstream and downstream relationship of the process are represented as a distance measurement function with a nonlinear weight: ; wherein, is the weighted distance between the multi-dimensional feature vector and the multi-dimensional feature vector , is the Euclidean distance between the multi-dimensional feature vector and the multi-dimensional feature vector , is the time proximity weight between the multi-dimensional feature vector and the multi-dimensional feature vector , is the process relationship weight between the multi-dimensional feature vector and the multi-dimensional feature vector .

[0010] Preferably, the time proximity weight between the multi-dimensional feature vector and the multi-dimensional feature vector is calculated by the following formula: ; wherein, is a preset time decay coefficient,​ a multi-dimensional feature vector a multi-dimensional feature vector a timestamp difference value; a multi-dimensional feature vector a multi-dimensional feature vector a process relationship weight is calculated by the following way: a multi-dimensional feature vector the process to which the multi-dimensional feature vector belongs a multi-dimensional feature vector the process to which the multi-dimensional feature vector belongs when and are the same process, is 1.0; when and are adjacent processes, is 1.5; when and are non-adjacent processes, is 3.0.

[0011] Preferably, the clustering of the multi-dimensional feature vectors into a plurality of process feature data layers comprises: setting a cutoff distance as the median of the distance value set between all data points; for each multi-dimensional feature vector , calculating the local density of ; the local density is the number of vectors in all other multi-dimensional feature vectors with a distance less than from ; calculating the minimum distance of the multi-dimensional feature vector ; the minimum distance is the minimum distance between and all multi-dimensional feature vectors with a local density greater than ; selecting the top 5% of multi-dimensional feature vectors in the product of and as the clustering centers, and assigning the remaining multi-dimensional feature vectors to the process feature data layer to which the nearest clustering center belongs.

[0012] Preferably, the calculation of the cross-layer transfer entropy between any two process feature data layers comprises: The two process characteristic data layers are source data layer and target data layer respectively, data of the source data layer and the target data layer are formed into two continuous time sequences by extracting the mean value of a preset physical property of the data in each time step in the layer; the two continuous time sequences are binned and discretized to obtain a discrete time sequence of the source data layer and a discrete time sequence of the target data layer The cross-layer transfer entropy from X to Y is calculated according to the following formula, ; Wherein, is the cross-layer transfer entropy from X to Y, the joint probability , the conditional probability and the conditional probability are obtained by frequency statistics on historical production data.

[0013] Preferably, the preset threshold is calculated by the following method: Extract the data of the last 30 normal production batches in the historical database; calculate the cross-layer transfer entropy between all data layers to form a cross-layer transfer entropy sample set; arrange all cross-layer transfer entropies greater than zero in the cross-layer transfer entropy sample set in ascending order; select the cross-layer transfer entropy at the 95th percentile after arrangement as the preset threshold.

[0014] Preferably, the determination threshold of the positive correlation between the historical edge weight of the directed dependency edge and the directed dependency edge includes: Extract the historical edge weight W of the directed dependency edge, which is the cross-layer transfer entropy calculated when constructing the acyclic directed graph; The determination threshold is calculated by the following formula: ; wherein, is the determination threshold, is a preset proportion coefficient, , is the historical edge weight.

[0015] Preferably, the calculation of the mutual information between real-time data in the process characteristic data layer includes: The real-time production data corresponding to the two process characteristic data layers connected by the directed dependency edge are formed into two real-time data sequences A and B by extracting the mean value of a preset physical property of each process characteristic data layer in each time step; the mutual information between A and B is calculated according to the following formula, ; Wherein, ​The mutual information between A and B, the joint probability p(a, b) is obtained by two-dimensional histogram statistics on the real-time data sequence A and B, the marginal probability p(a) is obtained by one-dimensional histogram statistics on the real-time data sequence A, and the marginal probability p(b) is obtained by one-dimensional histogram statistics on the real-time data sequence B.

[0016] In a second aspect, the circuit board production data management system based on big data provided by the present application comprises a memory and a processor, the memory stores computer instructions, and the processor executes the computer instructions to realize the circuit board production data management method based on big data.

[0017] The present application has the advantages that: the present application can sort the heterogeneous original production data into a structured process characteristic data layer according to the internal relationship of the original production data in the process flow and time. The non-cyclic directed graph constructed by calculating the cross-layer transfer entropy can convert the fuzzy and implicit influence relationship between processes into a causal transfer network. In abnormality determination, the change of data correlation strength on the key causal path is monitored to identify abnormalities, which can monitor early risks of weakening of the internal correlation of each parameter value itself within the specification range, and realize early warning of potential quality problems. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the circuit board production data management method based on big data provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0019] The embodiments of the present application will be described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0020] As shown in the figure, the embodiment of the circuit board production data management method based on big data provided by the present application comprises the following steps: Figure 1 S1, obtaining original production data in the circuit board production process, and constructing each data point into a multi-dimensional feature vector according to the time stamp, the belonging process and the data physical property.

[0021] ​Raw production data such as temperature, pressure, current, speed, chemical concentration, etc. are collected from manufacturing execution system (MES), programmable logic controller (PLC) and various sensors in real time, and each data is associated with a unique production batch number, equipment number, process code and collection timestamp. For example, the data point of 30000 rpm of the main shaft collected at 10:01:30 am in the drilling process is represented as a four-dimensional feature vector containing the timestamp, process code 1, attribute code 5 and value 30000.

[0022] S2, the density peak-based clustering algorithm is used, and the time proximity and process upstream and downstream relationship are represented as a distance metric function with nonlinear weights to cluster the multi-dimensional feature vectors into multiple process feature data layers.

[0023] Specifically, for any two multi-dimensional feature vectors, the Euclidean distance in time, process and physical attribute value is calculated. The distance metric function with nonlinear weights is used to correct the Euclidean distance. The distance metric function with nonlinear weights includes a time proximity weight function and a process relationship weight function. The time proximity weight function makes the weight between vectors with a time difference of seconds close to 1, and the weight with a time difference of hours increases exponentially. The process relationship weight function makes the weight between vectors belonging to the same process or adjacent processes smaller, and the weight between vectors belonging to processes far apart in the process flow increases. The density peak-based clustering algorithm is used to calculate the local density of the multi-dimensional feature vector, i.e. the number of neighbors with a corrected distance less than a certain cutoff distance, and the minimum distance to the high-density point. The multi-dimensional feature vector with larger local density and minimum distance is selected as the clustering center, and the remaining multi-dimensional feature vectors are attributed to the process feature data layer to which the nearest clustering center belongs.

[0024] S3, the cross-layer transfer entropy between any two process feature data layers is calculated, and when the cross-layer transfer entropy is greater than a preset threshold and the direction is consistent with the process flow, a directed dependence relationship edge is established between the two process feature data layers, and the cross-layer transfer entropy is used as the edge weight to construct a non-cyclic directed graph representing the internal causal association of data.

[0025] Specifically, first, all data points contained in any two process feature data layers A and B are extracted and arranged into two continuous time series in chronological order. Then, the cross-layer transfer entropy from A to B is calculated, which quantifies the degree of reduction of the historical information of data layer A on the uncertainty of the future state of data layer B. For example, the cross-layer transfer entropy of the lamination process temperature data layer A to the subsequent curing process temperature data layer B is calculated. At the same time, the system determines whether A is the upstream process of B according to the preset process sequence. Only when the calculated value of the cross-layer transfer entropy is greater than a threshold value set based on historical experience and data layer A is indeed upstream of B in the process sequence, the system will establish a directed dependency edge from A to B between the nodes representing A and B, and the cross-layer transfer entropy value will be used as the weight of the edge. By traversing all data layer pairs and repeating the above operation, a non-cyclic directed graph with process feature data layers as nodes and data intrinsic causal relationship as edges can be constructed.

[0026] S4, production data of the to-be-monitored circuit board production batch is extracted and mapped to the corresponding process feature data layer, for any pair of process feature data layers connected by a directed dependency edge in the non-cyclic directed graph, the mutual information between real-time data in the process feature data layer is calculated, and a determination threshold positively correlated with the historical edge weight of the directed dependency edge is calculated, and when the mutual information is lower than the determination threshold, it is determined that the data correlation of the to-be-monitored circuit board production batch is abnormal.

[0027] Specifically, after the start of a new production batch, the system calculates the modified distance between the generated real-time data vector and all cluster centers, and assigns it to the process feature data layer represented by the nearest cluster center. For any pair of data layers A and B connected by a dependency edge in the non-cyclic directed graph, the system sets a determination threshold positively correlated with its historical edge weight. Subsequently, the system collects all real-time data mapped to layers A and B in the current batch, and calculates the mutual information between these data. If the calculated mutual information value is lower than the set determination threshold, it is determined that the data correlation strength between A and B processes of the production batch is weaker than the historical normal level, there is a potential process correlation anomaly, and a warning is issued.

[0028] In an optional embodiment, the construction of each data point into a multi-dimensional feature vector according to the timestamp, the process to which it belongs, and the data physical property includes: The timestamp is converted into cumulative seconds since the start of the production batch, the process to which it belongs is converted into a binary vector using one-hot encoding to obtain a process vector, the data physical property is standardized by Z-score to obtain a standardized physical property value, and the data physical property includes temperature, pressure, and chemical concentration. The cumulative seconds, the process vector, and the standardized physical property value are concatenated in order to obtain the multi-dimensional feature vector.

[0029] Specifically, the construction process of the multi-dimensional feature vector is illustrated in detail with a single data point as an example. First, the time stamp and process information are processed: the time stamp is converted into the cumulative seconds since the start of the production batch as the first numerical element of the vector (e.g., 100 s). The process to which the data point belongs (e.g., exposure) is converted into a binary vector using one-hot encoding. If the entire production process includes three processes of cleaning, exposure, and etching, the exposure process is represented as a three-dimensional vector (0, 1, 0) immediately after the cumulative seconds. Second, the physical properties are subjected to Z-score standardization: assuming that the original physical properties of the data point are: temperature 25℃, pressure 101 kPa, and chemical concentration 0.5 mol / L. Z-score standardization is performed using historical statistical data (e.g., temperature mean 23℃, standard deviation 2℃; pressure mean 100 kPa, standard deviation 0.5 kPa; concentration mean 0.4 mol / L, standard deviation 0.05 mol / L). The standardization results are: temperature 1.0; pressure 2.0; and concentration 2.0, respectively. Finally, the processed cumulative seconds, process vector, and standardized physical property values are concatenated in order to obtain the multi-dimensional feature vector (100, 0, 1, 0, 1.0, 2.0, 2.0).

[0030] In an optional embodiment, the distance metric function representing the time proximity and process upstream / downstream relationship as a nonlinear weight is: ; wherein, is the multi-dimensional feature vector is the weighted distance between the multi-dimensional feature vector and the multi-dimensional feature vector is the Euclidean distance between the multi-dimensional feature vector and the multi-dimensional feature vector is the time proximity weight of the multi-dimensional feature vector and the multi-dimensional feature vector is the process relationship weight of the multi-dimensional feature vector and the multi-dimensional feature vector . The multi-dimensional feature vector and the multi-dimensional feature vector

[0031] The time proximity weight of the multi-dimensional feature vector and the multi-dimensional feature vector is calculated by the following formula: ; wherein, is a preset time decay coefficient, is the time stamp difference between the multi-dimensional feature vector and the multi-dimensional feature vector . ​Multi-dimensional feature vector The multi-dimensional feature vector The process relationship weight is calculated by the following way: The multi-dimensional feature vector The process to which the multi-dimensional feature vector belongs is The multi-dimensional feature vector The multi-dimensional feature vector is the same process as , the process relationship weight is 1.0; when and are adjacent processes, the process relationship weight is 1.5; when and are non-adjacent processes, the process relationship weight is 3.0.

[0032] For example, the weighted distance between two multi-dimensional feature vectors and is calculated. Suppose the vector is generated at 100s in the exposure process, and the vector is generated at 120s in the adjacent etching process. The Euclidean distance d(i, j) between the two vectors is 5.8 by calculating the difference between the feature values. Then the Euclidean distance is adjusted by the time proximity weight and the process relationship weight. The time stamp difference is 20s, and if the preset time decay coefficient a is 0.01, the time proximity weight is about 0.82, which reflects that the data points are more closely related in time. Since the exposure and etching are adjacent processes in the process flow, the process relationship weight is set to 1.5, which is greater than the weight 1.0 of the same process and less than the weight 3.0 of the non-adjacent process, reflecting the correlation strength between processes. The weighted distance D(i, j) is calculated to be about 7.13, which comprehensively considers the difference between the data and their relative positions in time and process flow.

[0033] In an alternative embodiment, the clustering of the multi-dimensional feature vectors into a plurality of process feature data layers comprises: The cutoff distance is the median of the distance value set between all data points; for each multi-dimensional feature vector , the local density of is calculated; the local density is the number of vectors with a distance less than to in all other multi-dimensional feature vectors; The multi-dimensional feature vectors minimum distance The minimum distance For all local densities greater than In the multidimensional feature vector, with The minimum distance between them; Select and The top 5% of the multidimensional feature vectors in the product are used as cluster centers, and the remaining multidimensional feature vectors are assigned to the process feature data layer to which the nearest cluster center belongs.

[0034] For example, in a dataset containing 1000 feature vectors, calculating the weighted distance between all vector pairs will produce nearly half a million distance values. After sorting these distance values, the value at the middle position is taken as the cutoff distance. For example, 15.2. Then, calculate the local density of each vector (e.g., vector A) in the dataset. Local density The calculation method is as follows: count how many vectors have a distance to vector A less than 15.2. Assume there are 50 such vectors. Then calculate the local density of vector A. It is 50.

[0035] Then, calculate the minimum distance of vector A. The specific method is as follows: Among all vectors with a local density greater than 50, find the one closest to vector A and record this distance value, let's say 25.6. If there is no point with a higher density than vector A, then the minimum distance is the farthest distance between vector A and all other points. Then, multiply the local density by the minimum distance to obtain a decision value. Repeat this calculation for all 1000 vectors and sort them according to the decision values, selecting the top 5% of 50 vectors as cluster centers. For each of the remaining 950 vectors, calculate the corrected distance to these 50 cluster centers and then assign it to the process feature data layer represented by the nearest cluster center.

[0036] In an optional embodiment, calculating the cross-layer transfer entropy between any two process feature data layers includes: The two process feature data layers are the source data layer and the target data layer, respectively. The data from the source and target data layers are processed by extracting the mean of a preset physical attribute at each time step to form two continuous time series. These two continuous time series are then binned and discretized to obtain the discrete time series of the source data layer. and discrete time series of the target data layer Calculate the cross-layer transfer entropy from X to Y using the following formula. ; in, Joint probability Conditional probability Joint probability Conditional probability

[0037] For example, assume that the transfer entropy from the exposure data layer to the etching data layer needs to be calculated, and temperature is chosen as the key physical property. First, the average temperature values of the two data layers at each time step are extracted from the historical data, forming two continuous time series, denoted as sequence X (exposure layer) and sequence Y (etching layer). For example, sequence X is (25.1, 25.3, 25.2, …), and sequence Y is (35.5, 35.4, 35.6, …). Next, the two continuous temperature sequences are discretized by binning. The entire temperature range is divided into a limited number of discrete intervals, for example, 20-30°C is divided into state 1, and 30-40°C is divided into state 2. After discretization, the original continuous sequences are converted into discrete sequences composed of integer states. Then, by counting the frequencies of the discrete sequences in a large amount of historical data, the various probabilities required for the transfer entropy calculation can be calculated. For example, the frequency of the combination of the target layer Y being in state 2 at the next time step, the current time step being in state 2, and the source layer X being in state 1 at the current time step is counted, as well as the frequency of the combination of the target layer Y being in state 2 at the next time step and the current time step being in state 2. Finally, the probability values obtained by frequency counting are substituted into the transfer entropy calculation formula to sum up, and a numerical value quantifying the amount of information transfer from the exposure layer temperature change to the etching layer temperature change can be obtained.

[0038] In an optional embodiment, the preset threshold is calculated as follows: Extract the data of the last 30 normal production batches in the historical database; calculate the transfer entropy between all data layers to form a sample set of transfer entropy; arrange all transfer entropies greater than zero in the sample set in ascending order; and select the transfer entropy at the 95th percentile after arrangement as the preset threshold.

[0039] For example, first, the system filters the last 30 production batches with qualified product quality from the production history. For each of the 30 qualified batches, the system performs data stratification and computation of cross-stratum transfer entropy. Assuming that the data is stratified into 10 process characteristic data strata, then for each batch, the system needs to compute the transfer entropy values between 90 pairs of directed data strata. Subsequently, the system aggregates all the computation results from the 30 batches to obtain a sample set containing 2700 transfer entropy values. The system performs data cleaning and threshold determination on the sample set. Through data cleaning, the system removes all the transfer entropy values that are less than or equal to zero, because these values do not represent valid information flow. Assuming that after removing the values less than or equal to zero, there are 2500 positive values left, the system ranks the 2500 positive values in ascending order and determines the position of the 95th percentile (i.e., the 2375th position). The transfer entropy value at this position is selected as the preset threshold. In subsequent causal relationship modeling, only when the computed cross-stratum transfer entropy value exceeds the preset threshold, the system determines that there is a strong association that needs to establish a dependency relationship edge.

[0040] In an optional embodiment, the determining a threshold value positively correlated with the historical edge weight of the directed dependency relationship edge comprises: extracting the historical edge weight W of the directed dependency relationship edge, the historical edge weight being the cross-stratum transfer entropy computed when constructing the acyclic directed graph; the threshold value being computed by the following formula: wherein, is the threshold value, is a preset proportionality coefficient, , is the historical edge weight.

[0041] For example, when monitoring the production process in real time, for a known dependency relationship, such as a directed edge from the cleaning data stratum to the exposure data stratum, the system queries the historical weight W of the directed edge from the acyclic directed graph constructed based on historical normal data. The weight W is the cross-stratum transfer entropy computed in history, representing the average intensity of the influence of the cleaning process on the exposure process under normal production conditions. Subsequently, the system introduces a proportionality coefficient β to adjust the sensitivity of the monitoring. A larger β value means that the system will detect even a small decrease in information flow intensity. By multiplying the historical edge weight W by the proportionality coefficient β, the system can compute the threshold value. The threshold value is used to determine whether the real-time information intensity between the current cleaning stratum and the exposure stratum has abnormally decreased.

[0042] In an optional embodiment, the computing the mutual information between real-time data in the process characteristic data strata comprises: corresponding to two process feature data layers connected by a directed dependency edge, form two real-time data sequences A and B by extracting the mean value of a certain pre-set physical property of each process feature data layer at each time step; calculate the mutual information between A and B according to the following formula, ; wherein, is the mutual information between A and B, the joint probability p(a, b) is obtained by two-dimensional histogram statistics of real-time data sequences A and B, the marginal probability p(a) is obtained by one-dimensional histogram statistics of real-time data sequence A, and the marginal probability p(b) is obtained by one-dimensional histogram statistics of real-time data sequence B.

[0043] For example, to calculate the mutual information between the cleaning layer and the exposure layer in real time, the system collects data within a continuous time window (for example, 10 minutes). First, select pressure as the representative attribute of the cleaning layer and temperature as the representative attribute of the exposure layer. By calculating the mean value of the corresponding attributes in these two data layers at each time step (for example, every 5 seconds), two real-time data sequences with 120 data points each can be obtained, which are denoted as sequence A (pressure) and sequence B (temperature). Then, the histogram method is used to estimate the probability distribution: the pressure value range of sequence A is divided into 5 intervals, and the temperature value range of sequence B is divided into 4 intervals. By counting how many of the 120 data fall into each grid of the two-dimensional grid, the estimated value of the joint probability p(a, b) can be obtained. The frequencies of the data of sequences A and B falling into their respective one-dimensional intervals are counted respectively, and the estimated values of the marginal probabilities p(a) and p(b) are obtained. Finally, the probability values obtained by histogram statistics are substituted into the mutual information calculation formula. The sum of the calculation results of all 20 grids can obtain the mutual information between the cleaning layer pressure and the exposure layer temperature within the current 10 minutes, indicating the correlation strength of the two under the current production state.

[0044] The implementation principle of the circuit board production data management method based on big data of the embodiments of the present application is that the present application can integrate original production data with diverse sources and different structures, and according to the internal logical relationship thereof in the process flow and the time dimension, reconstruct it into a structured process feature data layer. In addition, by calculating the transfer entropy to construct a non-cyclic directed graph, the original vague and implicit mutual influence between processes can be effectively revealed and visualized as a clear causal relationship network. Moreover, in terms of anomaly detection, the present application identifies risks by monitoring the change in data correlation strength on the key causal path in real time. Even if all parameter readings are within the normal range, early signs of weakening correlation between them can be captured, thereby achieving early warning of potential quality problems.

[0045] ​The embodiment of the circuit board production data management system based on big data provided by the application comprises a memory and a processor, the memory stores computer instructions, and the processor executes the computer instructions to realize the circuit board production data management method based on big data in the above embodiment.

[0046] Although the embodiments of the application have been shown and described above, it should be understood by those skilled in the art that the above embodiments are exemplary and cannot be understood as limiting the application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the application.

Claims

1. A method for managing data of a circuit board production based on big data, characterized by, The method comprises the following steps: obtaining original production data in a circuit board production process, and constructing each data point into a multi-dimensional feature vector according to a time stamp, a process to which the data point belongs, and a physical attribute of the data point; using a clustering algorithm based on density peaks, and using a distance measurement function in which time proximity and upstream and downstream relationships in the process are represented as nonlinear weights to cluster the multi-dimensional feature vector into a plurality of process feature data layers; calculating cross-layer transfer entropy between any two process feature data layers, and when the cross-layer transfer entropy is greater than a preset threshold and the direction is consistent with the process flow, establishing a directed dependency relationship edge between the two process feature data layers, and using the cross-layer transfer entropy as an edge weight to construct a non-cyclic directed graph representing the intrinsic causal relationship of the data; extracting production data of a to-be-monitored circuit board production batch and mapping the data to corresponding process feature data layers, calculating mutual information between real-time data in the process feature data layers for any pair of process feature data layers connected by a directed dependency relationship edge in the non-cyclic directed graph, and calculating a judgment threshold positively correlated with a historical edge weight of the directed dependency relationship edge, when the mutual information is lower than the judgment threshold, determining that data correlation of the to-be-monitored circuit board production batch is abnormal.

2. The big data-based circuit board production data management method of claim 1, wherein, The construction of each data point into a multi-dimensional feature vector according to a time stamp, a process to which the data point belongs, and a physical attribute of the data point comprises: converting the time stamp into cumulative seconds from the start of the production batch, converting the process to which the data point belongs into a binary vector using one-hot encoding to obtain a process vector, performing Z-score standardization on the physical attribute of the data point to obtain a standardized physical attribute value, the physical attribute of the data point including temperature, pressure, and chemical concentration, and concatenating the cumulative seconds, the process vector, and the standardized physical attribute value in sequence to obtain the multi-dimensional feature vector.

3. The big data-based circuit board production data management method of claim 1, wherein, The distance measurement function in which time proximity and upstream and downstream relationships in the process are represented as nonlinear weights is: ； wherein, is a multi-dimensional feature vector is a multi-dimensional feature vector is a weighted distance between is a multi-dimensional feature vector is a Euclidean distance between is a multi-dimensional feature vector is a multi-dimensional feature vector is a temporal proximity weight of is a multi-dimensional feature vector is a multi-dimensional feature vector is a process relationship weight of is a multi-dimensional feature vector 4. The big data-based circuit board production data management method of claim 3, characterized by, Multi-dimensional feature vector Time proximity weight with multi-dimensional feature vector Time proximity weight with multi-dimensional feature vector is calculated by the following equation: ; wherein, is a preset time decay coefficient, is a multi-dimensional feature vector and the multi-dimensional feature vector timestamp difference; Multi-dimensional feature vector Process relationship weight with multi-dimensional feature vector Process relationship weight with multi-dimensional feature vector By the following way: multidimensional feature vector The procedure is , multidimensional feature vector The procedure is when and are the same procedure, is 1.0; when and are adjacent procedures, is 1.5; when and are non-adjacent procedures, is 3.

0.

5. The big data-based circuit board production data management method of claim 1, wherein, The clustering of the multi-dimensional feature vector into a plurality of process feature data layers comprises: Set cutoff distance The median of the set of distance values ​​between all pairs of data points; for each multidimensional feature vector ,calculate Local density The local density To be among all other multidimensional feature vectors Distance less than The number of vectors; Computing a multidimensional feature vector of minimum distance ; the minimum distance is the minimum distance between the multidimensional feature vector and where all local densities are greater than Selecting and The top 5% of the multi-dimensional feature vectors of the product are selected as the clustering centers, and the remaining multi-dimensional feature vectors are assigned to the process feature data layer to which the nearest clustering center belongs.

6. The big data-based circuit board production data management method of claim 1, wherein, The calculation of cross-layer transfer entropy between any two process feature data layers comprises: The two process feature data layers are respectively source data layer and target data layer, data of the source data layer and the target data layer are formed into two continuous time sequences by extracting mean values of a preset physical property of data in each time step in the layers; the two continuous time sequences are binned and discretized to obtain a discrete time sequence of the source data layer and a discrete time sequence of the target data layer The cross-layer transfer entropy from X to Y is calculated according to the following formula,​ ； wherein, Joint probability Conditional probability Joint probability are obtained by frequency statistics on historical production data.

7. The big data-based circuit board production data management method of claim 1, wherein, The preset threshold is calculated by the following method: extracting data of the last 30 normal production batches in a historical database, calculating cross-layer transfer entropy between all data layers to form a cross-layer transfer entropy sample set, and arranging all cross-layer transfer entropy greater than zero in the cross-layer transfer entropy sample set in ascending order; selecting cross-layer transfer entropy at the 95th percentile after arrangement as the preset threshold.

8. The big data-based circuit board production data management method of claim 1, wherein, The calculation of a judgment threshold positively correlated with a historical edge weight of a directed dependency relationship edge comprises: extracting a historical edge weight W of the directed dependency relationship edge, the historical edge weight being cross-layer transfer entropy calculated when the non-cyclic directed graph is constructed; The judgment threshold is calculated by the following formula: ; wherein, is a decision threshold, is a preset proportion coefficient, , is a historical edge weight.

9. The big data-based circuit board production data management method of claim 1, wherein, The calculation of mutual information between real-time data in a process feature data layer comprises: forming two real-time data sequences A and B by extracting the mean value of a certain preset physical attribute of each process feature data layer at each time step for real-time production data corresponding to two process feature data layers connected by a directed dependency relationship edge, and calculating mutual information between A and B according to the following formula, ； where, is the mutual information between A and B, the joint probability p(a, b) is obtained by two-dimensional histogram statistics on real-time data sequences A and B, the marginal probability p(a) is obtained by one-dimensional histogram statistics on real-time data sequence A, and the marginal probability p(b) is obtained by one-dimensional histogram statistics on real-time data sequence B.

10. A big data-based circuit board production data management system, characterized by, The application relates to a circuit board production data management method based on big data, comprising a memory and a processor, wherein the memory stores computer instructions, and the processor executes the computer instructions to realize the circuit board production data management method based on big data.

Images