Electronic product production risk assessment method and system based on artificial intelligence

Abstract

The application provides an electronic product production risk assessment method and system based on artificial intelligence, relates to the field of risk assessment, and solves the technical problems of fuzzy cause-effect relationship and insufficient response to distribution deviation during electronic product production risk assessment. The method first extracts material, process, equipment and environment heterogeneous elements of the production process, constructs an MPEE standard data set and a heterogeneous cause-effect diagram; collects real-time MPEE data through a fixed time sliding window to construct a sliding window data set, detects the distribution deviation degree of the data set from the standard data set by using a kernel function, sets a threshold value in combination with process stability and locates the deviation nodes, realizes dynamic updating of the heterogeneous cause-effect diagram, inputs the real-time data into a risk assessment model trained by a deep neural network to obtain a risk result, reversely traverses a cause-effect path based on the risk level and the heterogeneous cause-effect diagram, completes root cause positioning and impact degree quantification, and finally matches a process knowledge base strategy and selects an optimal execution strategy.

Description

Technical Field

[0001] This application relates to the field of risk assessment, and in particular to a method and system for risk assessment in the production of electronic products based on artificial intelligence. Background Technology

[0002] The electronics manufacturing industry is a technology-intensive and precision manufacturing industry. Its production process covers multiple processes such as solder paste printing, SMT reflow soldering, precision assembly, and electrical testing. The processes are highly coupled and require high precision. Moreover, the production process is dynamically affected by various heterogeneous factors such as material characteristics, equipment status, process parameters, and environmental conditions.

[0003] Current risk assessments in electronic product manufacturing primarily rely on human experience and traditional statistical analysis methods. However, the correlations between production factors are often simply judged as statistical correlations, failing to accurately identify causal relationships. This results in an inability to quickly pinpoint the root cause after a risk occurs, only addressing superficial anomalies. Consequently, risk recurrence rates are high, and the distributional deviation between real-time and standard data during production is not effectively detected and addressed. Risk assessments based on this deviation are distorted and fail to accurately reflect the actual risk status of production. Summary of the Invention

[0004] This application provides an artificial intelligence-based method and system for risk assessment in electronic product manufacturing, which solves the technical problems of ambiguous causal relationships and insufficient response to distribution deviations in the risk assessment of electronic product manufacturing in the prior art.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides an AI-based method for assessing risks in electronic product manufacturing, including: To acquire heterogeneous elements in the electronic product manufacturing process and construct a heterogeneous cause-effect graph and standard MPEE for electronic product manufacturing processes; Real-time MPEE data of the production process is collected using a fixed-time sliding window, and a sliding window dataset D is constructed. t and the standard dataset D0; Based on the sliding window dataset D t The distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph; The data corresponding to the nodes of real-time MPEE are input into the risk assessment model to obtain the risk result; wherein, the risk assessment model is trained based on a deep neural network model; the risk result includes risk level and risk node; Root cause localization and causal contribution quantification are performed based on risk level and heterogeneous causal graph to obtain root cause nodes and their degree of influence; the optimal execution strategy is determined based on the root cause nodes and their degree of influence.

[0006] Based on the above technical solutions, the risks in the production process of electronic products are continuously and dynamically assessed, and the production process is continuously optimized through cyclical feedback. This can further improve the efficiency and quality of the electronic product production process. From the construction of heterogeneous cause-effect graphs, real-time data acquisition, distribution offset detection and cause-effect graph updates, to risk assessment, root cause localization, and optimal strategy execution, an end-to-end production risk management process is formed. Furthermore, through data feedback in the process, the heterogeneous cause-effect graphs and risk assessment models are continuously optimized, realizing cyclical management of production risks and continuously improving the process stability and product yield of electronic product production.

[0007] In conjunction with the first aspect above, in one possible implementation, the method for constructing the heterogeneous causal graph includes: Extract the MPEE of each process in the production of electronic products; where MPEE includes material node M, process node P, equipment node E, and environment node En. MPEE elements are defined as nodes in a causal graph. Directed causal edges between causal nodes are constructed using a constrained causal model. The weights of the directed causal edges are quantified. The causal graph nodes, directed causal edges, and their corresponding weights are then integrated to obtain a preliminary causal graph. The preliminary causal graph is validated, and the validated preliminary causal graph is marked as a heterogeneous causal graph.

[0008] In conjunction with the first aspect above, in one possible implementation, the quantification method of the weight of the directed causal edge includes: Retrieve the cause node X and effect node Y from the cause-effect graph; Determine whether cause node X and effect node Y are consecutive nodes; No, the average causal effect is calculated using the potential outcome framework and labeled as the weight of the directed causal edge; the average causal effect is calculated as follows: ; The average causal effect of node X on node Y; do(X=1) is the action that forces node X to be in an abnormal state; This represents the potential outcome of Y when X=1; Let X be the potential outcome of Y when X=0; do(X=0) is the dynamic that forces node X to the normal state; E[] is the mathematical expectation; Yes, based on the process threshold, node X is divided into a normal domain X0 and an abnormal domain X1; then, through calculation... Calculate the average causal effect and label it as the weight of the directed causal edge; where, Let X be the sample size when node X is in the outlier region X1. Let X be the sample size when node X is in the normal domain X0. Let Y be the observed value corresponding to the i-th X anomaly sample. Let Y be the observed value corresponding to the j-th normal X sample.

[0009] It should be noted that if the cause node X is a discrete node and the result node Y is a continuous node, then the discrete cause node X needs to be binarized according to whether it is normal or abnormal before using the calculation formula. Calculate the corresponding weights; When determining whether cause node X and effect node Y are consecutive nodes, a preset mapping table can be used for this determination. The specific process for determining whether a node is consecutive can be followed: Determine whether the node can be continuously adjusted or continuously changed in the process it is in; if the process is classified, graded, counted, or not continuously adjustable, it is directly determined to be a discrete node; if the process can be continuously adjusted or continuously changed, and the value is in a range, it is determined to be a continuous node.

[0010] By extracting four types of heterogeneous nodes in MPEE, constructing directed causal edges and verifying them, the causal graph can accurately reflect the actual causal relationships of various elements in the production of electronic products, avoiding confusion of causal relationships caused by simple statistical correlations. This lays a reliable model foundation for subsequent root cause localization and causal contribution quantification, and improves the scientific nature and practical relevance of causal analysis.

[0011] In conjunction with the first aspect above, in one possible implementation, the method for updating the heterogeneous causal graph includes: The kernel-based MMD algorithm is used to compute the detection of the sliding window dataset D. t The distribution difference between the dataset and the standard dataset D0; the formula for calculating the distribution difference is as follows: ; The sample size of the standard dataset D0, Sliding window dataset D t Sample size; Let i be the MPEE feature vector of the i-th and j-th samples in the standard dataset D0; For the sliding window dataset D t The MPEE feature vectors of the m-th and n-th samples in the dataset; is a Gaussian kernel function used to map high-dimensional eigenvectors to a reproducing kernel Hilbert space; Set the distribution offset threshold based on the stability of electronic product manufacturing processes. ; Distribution difference and distribution offset threshold Compare; when the distribution difference is greater than the distribution offset threshold If the distribution differences are analyzed using the feature layer MMD decomposition method to determine the offset nodes, the heterogeneous causal graph is updated based on the offset nodes; otherwise, the heterogeneous causal graph is saved.

[0012] In conjunction with the first aspect above, in one possible implementation, the step of setting a distribution offset threshold based on the stability of the electronic product manufacturing process is... ,include: Collect historical data on the electronic product manufacturing process where no anomalies occurred, and construct a benchmark dataset D. 0-m ; Using a sliding window on the benchmark dataset D 0-m Self-comparison was performed to obtain MMD with normal fluctuations. 2 Statistical characteristics; Based on MMD 2 Statistical characteristics are used to calculate the distribution offset threshold; where the calculation formula is: ; For MMD 2 The mean of a statistical characteristic; This is the process capability coefficient. ; This is the quantile correction factor. For MMD 2 Standard deviation of statistical characteristics This represents the sensitivity coefficient to process risks.

[0013] It should be noted that, CPK is a process capability index; the more stable the process, the larger the CPK. The smaller, The smaller the value, the more sensitive it is to offset.

[0014] In conjunction with the first aspect above, in one possible implementation, the method of determining the offset node includes: Extract the MPEE nodes from the production process to obtain a node list; The proportion of each node's ATE value is marked as the node weight, and a specific offset threshold is calculated based on the node weight and the distribution offset threshold; wherein, the formula for calculating the specific offset threshold is: ;in, Let be the node weight of the k-th node; This is the exclusive offset threshold for the k-th node; The distribution offset of each node is calculated by kernelizing the MMD formula; The distribution offset of each node is compared with the exclusive offset threshold, and the nodes whose distribution offset is greater than the exclusive offset threshold are marked as offset nodes.

[0015] It should be noted that the formula for calculating the degree of distribution shift is: In the formula, k is the node number. The sample size of the standard dataset D0, Sliding window dataset D tSample size The degree of distribution offset for the k-th node alone. For the i-th sample of the k-th node in the standard dataset, This refers to the m-th sample of the k-th node in the sliding window dataset. If multiple nodes simultaneously satisfy Then according to The ratios are sorted from high to low, and the nodes corresponding to this sorting are updated sequentially.

[0016] In conjunction with the first aspect above, in one possible implementation, the root cause localization and causal contribution quantification based on risk outcomes and heterogeneous causal graphs includes: Taking the risk node in the risk outcome as the endpoint, and traversing in the opposite direction of the directed causal edge in the heterogeneous causal graph, we obtain several upstream causal paths and upstream MPEE nodes leading to the risk node. Traverse the source nodes in several upstream causal paths; Extract the ATE value from the upstream causal path and determine the number of steps and chain weights in the upstream causal path; calculate the formula... Calculate the influence coefficient (TCC) of all source nodes; where n is the number of causal edges in the l-th indirect causal chain. Let L be the chain weight, and L be the number of indirect causal chains from X to Y. Let be the ATE value of the i-th causal edge in the l-th indirect causal chain; Traverse the source nodes corresponding to the maximum value of the influence level TCC and mark them as root cause nodes.

[0017] In conjunction with the first aspect above, in one possible implementation, determining the optimal execution strategy based on the root cause node and the degree of influence includes: Extract the manufacturing process knowledge base of electronic products; the manufacturing process knowledge base includes risk suppression strategies for various MPEE node anomalies. Candidate suppression strategies are matched based on root cause nodes; the optimal strategy is selected based on the strategy utility value of each candidate suppression strategy, and the optimal strategy is executed.

[0018] Furthermore, the strategy utility value can be calculated by weighted summation based on the risk suppression result, intervention cost, and production interruption impact. Among multiple candidate suppression strategies, the strategy with the highest strategy utility value is the optimal strategy. Specifically, the risk suppression result can be calculated based on the ratio of the decrease in risk level after strategy implementation to the current risk level; the larger the value, the better the suppression effect. The intervention cost is calculated based on the ratio of the actual cost of implementing the strategy to the maximum acceptable cost of a single intervention for the enterprise; the smaller the value, the lower the cost. The production interruption impact can be calculated based on the production interruption time caused by strategy implementation to the maximum acceptable interruption time of a single intervention for the enterprise; the smaller the value, the smaller the impact on production.

[0019] In conjunction with the first aspect above, in one possible implementation, the method for constructing the risk assessment model includes: Acquire historical risk data for a number of electronic products and production lines; divide training data into task sets; wherein each task set corresponds to a heterogeneous cause-effect graph and risk labeling data for an electronic product production line; A training model is constructed based on a deep neural network, with heterogeneous causal graph nodes as inputs and risk levels and probabilities as outputs. By performing multi-task gradient updates on the model to be trained using training data, a risk assessment model is obtained.

[0020] Secondly, this application provides an artificial intelligence-based electronic product manufacturing risk assessment system, the device comprising: a preprocessing module and an assessment processing module; The preprocessing module is used to acquire heterogeneous elements of the electronic product manufacturing process and construct a heterogeneous cause-effect graph and a standard MPEE for electronic product production processes; it also collects real-time MPEE of the production process based on a fixed-time sliding window and constructs a sliding window dataset D. t and the standard dataset D0; based on the sliding window dataset D t The distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph; The assessment and processing module is used to input the data corresponding to the nodes of real-time MPEE into the risk assessment model to obtain the risk results; wherein, the risk assessment model is trained based on a deep neural network model; root cause localization and causal contribution quantification are performed according to the risk level and heterogeneous causal graph to obtain the root cause nodes and the degree of influence.

[0021] This application provides an AI-based method and system for risk assessment in electronic product manufacturing. It traverses the reverse causal path of risk nodes using a heterogeneous causal graph, quantifies the causal contribution of root cause nodes using TCC values, and transforms empirical root cause judgment into quantitative root cause localization. This achieves a precise connection from risk discovery to root cause identification, significantly shortening root cause localization time and improving risk handling efficiency. Furthermore, from heterogeneous causal graph construction, real-time data acquisition, distribution offset detection and causal graph updates, to risk assessment, root cause localization, and optimal strategy execution, an end-to-end production risk management process is formed. The process continuously optimizes the heterogeneous causal graph and risk assessment model through data feedback, achieving cyclical management of production risks and continuously improving the process stability and product yield of electronic product manufacturing.

[0022] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0023] Figure 1 A system architecture diagram of an AI-based electronic product manufacturing risk assessment system provided in this application embodiment; Figure 2 A flowchart illustrating the AI-based risk assessment method for electronic product manufacturing provided in this application embodiment; Figure 3 A flowchart illustrating step S201 provided in an embodiment of this application; Figure 4 This is a schematic flowchart illustrating the method for updating a heterogeneous causal graph provided in an embodiment of this application. Detailed Implementation

[0024] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0025] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0026] The AI-based risk assessment method for electronic product manufacturing provided in this application can be applied to, for example... Figure 1 In the AI-based electronic product manufacturing risk assessment system 100 shown, such as Figure 1 As shown, the system includes: a data acquisition device 101, an industrial server 102, and an execution control device 103.

[0027] Among them, the data acquisition device 101 is used to acquire heterogeneous elements of the electronic product manufacturing process; and to acquire real-time MPEE of the production process based on a fixed-time sliding window.

[0028] Industrial server 102 is used to construct heterogeneous causal graphs and sliding window datasets D. t The system uses standard dataset D0 and standard MPEE for electronic product manufacturing processes; it inputs the data corresponding to the nodes of real-time MPEE into the risk assessment model to obtain risk results; it performs root cause localization and causal contribution quantification based on risk level and heterogeneous causal graph to obtain root cause nodes and degree of impact.

[0029] The execution control device 103 is used to determine the optimal execution strategy based on the root cause node and the degree of impact.

[0030] To address the technical problems of ambiguous causal relationships and insufficient response to distribution offsets in the risk assessment of electronic product manufacturing in the prior art, this application provides an artificial intelligence-based method for risk assessment of electronic product manufacturing. The method includes: acquiring heterogeneous elements of the electronic product manufacturing process and constructing a heterogeneous causal graph, a standard dataset D0, and a standard MPEE for electronic product manufacturing processes. Real-time MPEE data of the production process is collected using a fixed-time sliding window, and a sliding window dataset D is constructed. t ; Based on the sliding window dataset D t The distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph; The data corresponding to the nodes of real-time MPEE are input into the risk assessment model to obtain the risk result; wherein, the risk assessment model is trained based on a deep neural network model; the risk result includes risk level and risk node; Root cause localization and causal contribution quantification are performed based on risk level and heterogeneous causal graph to obtain root cause nodes and their degree of impact. The optimal execution strategy is determined based on the root cause nodes and their degree of impact. Based on this, the risks in the electronic product manufacturing process are continuously and dynamically assessed, and the production process is continuously optimized through cyclical feedback, which can further improve the efficiency and quality of electronic product manufacturing.

[0031] like Figure 2 As shown, this embodiment uses a mobile phone motherboard SMT reflow soldering process line as the application scenario. The core quality risk node of this process line is solder joint cold solder joints. The involved MPEE heterogeneous elements include: material node M (solder paste viscosity), process node P (soldering temperature, conveyor belt speed), equipment node E (heating tube power, pick-and-place machine speed), and environmental node En (workshop humidity, workshop temperature). The artificial intelligence-based electronic product manufacturing risk assessment method provided in this application embodiment specifically includes the following steps: S201. Obtain heterogeneous elements of the electronic product manufacturing process and construct a heterogeneous cause-effect graph, standard dataset D0, and standard MPEE for electronic product manufacturing processes.

[0032] MPEE includes material node M, process node P, equipment node E, and environment node En.

[0033] For example, the MPEE heterogeneous elements of an SMT reflow soldering process line are: material node M (solder paste viscosity), process node P (soldering temperature, conveyor belt speed), equipment node E (heating tube power, pick-and-place machine speed), and environment node En (workshop humidity, workshop temperature). The process attribute determination results for each node are as follows:

[0034] Continuous nodes: solder paste viscosity, soldering temperature, heating tube power, workshop humidity, workshop temperature, conveyor belt speed; since all of the above nodes can be continuously adjusted and their values ​​are within a range, they are all continuous nodes.

[0035] Discrete node: The pick-and-place machine speed is adjustable in only 3 levels, which is a graded attribute. Therefore, the pick-and-place machine speed cannot be changed continuously and gradually. In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S201 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Obtain the heterogeneous elements of the electronic product manufacturing process and construct a standard MPEE for electronic product production processes.

[0036] S302. Construct a heterogeneous causal graph.

[0037] In some implementations, methods for constructing heterogeneous cause-effect graphs include: Extract the MPEE of each process in the production of electronic products; where MPEE includes material node M, process node P, equipment node E, and environment node En. MPEE elements are defined as nodes in a causal graph. Directed causal edges between causal nodes are constructed using a constrained causal model. The weights of the directed causal edges are quantified. The causal graph nodes, directed causal edges, and their corresponding weights are then integrated to obtain a preliminary causal graph. The preliminary causal graph is validated, and the validated preliminary causal graph is marked as a heterogeneous causal graph.

[0038] It should be noted that the preliminary cause-effect graph was manually verified by three senior engineers in the SMT process field to confirm the rationality of the causal edges. Verification rule: If P(Y=risk|do(X=abnormal)) P(Y=risk|do(X=normal)): Anomalies in X are a positive cause of risks in Y (in most scenarios in the production of electronic products); If P(Y=risk|do(X=abnormal)) P(Y=Risk|do(X=Normal)): Anomalies in X are a negative cause of risk in Y (rarely occurs in industrial scenarios and require manual process verification). If P(Y=risk|do(X=abnormal))=P(Y=risk|do(X=normal)): X and Y have no causal relationship, and may only have a statistical association; The preliminary cause-effect graph can be verified by experts in the field. For example, invalid cause-effect edges of workshop temperature → solder paste viscosity (ATE=0.20, P(Y=risk|do(X=abnormal))≈P(Y=risk|do(X=normal))) are removed, and the remaining core cause-effect edges are retained to form a heterogeneous cause-effect graph.

[0039] In further ways of implementation, Methods for quantifying the weights of directed causal edges include: Retrieve the cause node X and effect node Y from the cause-effect graph; Determine whether cause node X and effect node Y are consecutive nodes; No, the average causal effect is calculated using the potential outcome framework and labeled as the weight of the directed causal edge; the average causal effect is calculated as follows: ; The average causal effect of node X on node Y; do(X=1) is the action that forces node X to be in an abnormal state; This represents the potential outcome of Y when X=1; Let X be the potential outcome of Y when X=0; do(X=0) is the dynamic that forces node X to the normal state; E[] is the mathematical expectation; Yes, based on the process threshold, node X is divided into a normal domain X0 and an abnormal domain X1; then, through calculation... Calculate the average causal effect and label it as the weight of the directed causal edge; where, Let X be the sample size when node X is in the outlier region X1. Let X be the sample size when node X is in the normal domain X0. Let Y be the observed value corresponding to the i-th X anomaly sample. Let Y be the observed value corresponding to the j-th normal X sample.

[0040] It should be noted that, This indicates the degree of causal influence of cause node X on effect node Y.

[0041] For example, for continuous cause-effect nodes, such as workshop humidity → solder paste viscosity, and heating element power → soldering temperature, the cause node is divided into a normal domain X0 and an abnormal domain X1 according to process thresholds. The normal domain X0 for workshop humidity is 30%RH-70%RH, and the abnormal domain X1 is <30%RH or >70%RH. The normal domain X0 for heating element power is 80W-90W, and the abnormal domain X1 is <80W or >90W. This is achieved through the formula... Calculate the ATE value and find that the ATE value for workshop humidity → solder paste viscosity is 0.85, and the ATE value for heating tube power → soldering temperature is 0.95. For discrete cause nodes and continuous effect nodes, the formula is used... Calculate the ATE value and find that the ATE value for the pick-and-place machine speed → soldering temperature is 0.15; Based on process correlations and ATE values, invalid causal edges with ATE values ​​close to 0 are removed (e.g., pick-and-place machine speed → solder joint cold solder joint, ATE=0.02). A preliminary causal graph is constructed, with the core directed causal edges being: workshop humidity → solder paste viscosity (ATE=0.85), solder paste viscosity → solder joint cold solder joint (ATE=0.90), heating element power → soldering temperature (ATE=0.95), soldering temperature → solder joint cold solder joint (ATE=0.85), and workshop temperature → solder paste viscosity (ATE=0.20).

[0042] S303. Construct the standard dataset D0.

[0043] In some implementations, MPEE data is collected during the stage when the SMT reflow soldering process line is operating without abnormalities and the product yield is ≥99.5%. A total of 5,000 samples are collected, covering normal production conditions of different shifts and different solder paste batches, and a standard dataset D0 is constructed. The data of each node meets the normal process threshold requirements.

[0044] S202. Real-time MPEE of production process data is collected using a fixed-time sliding window, and a sliding window dataset D is constructed. t .

[0045] In some implementations, a fixed time window of 5 minutes is set to collect real-time data from the MPEE node of the SMT reflow soldering process line. Data is collected every 5 minutes, with 20 samples collected each time, to construct a sliding window dataset D. t D t The collected metrics are completely consistent with D0, ensuring data comparability.

[0046] S203, Based on the sliding window dataset D t The degree of distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph.

[0047] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above-mentioned method for updating the heterogeneous cause-effect graph can be specifically implemented through the following steps S401, S402, and S403, which are explained in detail below: S401. Calculate the detection sliding window dataset D using kernel-optimized MMD. t The distribution difference between the standard dataset D0 and the standard dataset D0.

[0048] The formula for calculating the distribution difference is as follows: ; The sample size of the standard dataset D0, Sliding window dataset D t Sample size; Let i be the MPEE feature vector of the i-th and j-th samples in the standard dataset D0; For the sliding window dataset D t The MPEE feature vectors of the m-th and n-th samples in the dataset; is a Gaussian kernel function used to map high-dimensional eigenvectors to a reproducing kernel Hilbert space.

[0049] In this embodiment, the Gaussian kernel function Simplified calculation formula , (where N0 is the standard deviation of the baseline dataset for each node), substitute into D0 (N0=5000) and D t (N) t From the sample data of (=20), the following calculations were performed. .

[0050] S402. Set the distribution offset threshold according to the stability of electronic product manufacturing process. .

[0051] In some implementations, a distribution offset threshold is set based on the stability of the electronic product manufacturing process. ,include: Collect historical data on the electronic product manufacturing process where no anomalies occurred, and construct a benchmark dataset D. 0-m ; Using a sliding window on the benchmark dataset D 0-m Perform self-comparison (i.e., perform a benchmark dataset D) 0-m Randomly split into multiple subsets D 0-m1 D 0-m2 ...mutual calculation of MMD 2 ), to obtain MMD with normal fluctuations 2 Statistical characteristics; Based on MMD 2 Statistical characteristics are used to calculate the distribution offset threshold; where the calculation formula is: ; For MMD 2 The mean of a statistical characteristic; This is the process capability coefficient. ; This is the quantile correction factor. For MMD 2 Standard deviation of statistical characteristics This represents the sensitivity coefficient to process risks.

[0052] It should be noted that, CPK is a process capability index; the more stable the process, the larger the CPK. The smaller, The smaller the value, the more sensitive it is to offset; the value range is 0.2-1.0. A fixed value of 1.96 is used, corresponding to the 95th percentile in statistics, to ensure that the threshold covers 95% of normal fluctuations, while the process risk sensitivity coefficient... This is based on a pre-defined process precision; for example, the process risk sensitivity coefficient. Refer to Table 1: Table 1

[0053] In this embodiment, the SMT reflow soldering process belongs to the G1 high-precision process. The distribution offset threshold is calculated according to the following steps, and historical data without process anomalies are collected to construct a benchmark dataset D. 0-m MMD with normal fluctuations was obtained through self-comparison using the sliding window method. 2 Statistical characteristics: , The process has a CPK of 1.33 (extremely stable), therefore... The quantile correction factor α is fixed at 1.96, and the risk sensitivity coefficient of process G1 is... =1.2, substitute into the formula =0.03564.

[0054] S403. Compare the distribution difference with the distribution offset threshold τ. If the distribution difference is greater than the distribution offset threshold τ, analyze the distribution difference using the feature layer MMD decomposition method to determine the offset node. Update the heterogeneous causal graph based on the offset node. Otherwise, save the heterogeneous causal graph.

[0055] In this embodiment, the distribution difference and the offset threshold are compared. >0.03564; determined to be a significant distribution offset, initiate feature layer MMD decomposition to locate the offset node. It should be noted that the local update method of the heterogeneous causal graph is as follows: for the offset node, the PC algorithm is used to re-identify its causal edges and the weights of the causal edges, remove invalid causal edges, add new causal edges, and correct the causal strength.

[0056] Furthermore, the methods for determining offset nodes include: Extract the MPEE nodes from the production process to obtain a node list; The proportion of each node's ATE value is marked as the node weight, and a specific offset threshold is calculated based on the node weight and the distribution offset threshold; wherein, the formula for calculating the specific offset threshold is: ;in, Let be the node weight of the k-th node; This is the exclusive offset threshold for the k-th node; The distribution offset of each node is calculated by kernelizing the MMD formula; The distribution offset of each node is compared with the exclusive offset threshold, and the nodes whose distribution offset is greater than the exclusive offset threshold are marked as offset nodes.

[0057] It should be noted that the formula for calculating the degree of distribution shift is: In the formula, The degree of distribution offset for the k-th node alone. This refers to the i-th sample of the k-th node in the standard dataset (such as the normal value of solder paste viscosity). For the m-th sample of the k-th node in the sliding window dataset (e.g., the latest value of solder paste viscosity); If multiple nodes simultaneously satisfy Then according to The ratios are sorted from high to low, and the nodes corresponding to this sorting are updated sequentially.

[0058] For example, extract the core MPEE node list for the process: workshop humidity, solder paste viscosity, heating element power, and soldering temperature, and calculate the node weight ω according to the proportion of ATE value. k ω1 (workshop humidity) = 0.2, ω2 (solder paste viscosity) = 0.3, ω3 (heating tube power) = 0.15, ω4 (soldering temperature) = 0.35, and ∑ω k =1; Calculate the node-specific offset threshold. k = ×ω k : 1 = 0.007128 2 = 0.010692 3 = 0.005346 4 = 0.012474; based on single-node MMD 2 The formula calculates the degree of distribution offset of each node. , , , Solder paste viscosity only The solder paste viscosity was determined to be the only offset node. For the solder paste viscosity of the offset nodes, the PC algorithm is used to re-identify their causal edges and weights: The ATE value of solder paste viscosity due to workshop humidity was recalculated, and the new ATE value was 0.92. Recalculate the solder paste viscosity → ATE value for poor solder joint, and get a new ATE = 0.95; Invalid causal edges are removed, causal strength is adjusted, and a local update of the heterogeneous causal graph is completed, while the causal edges of the remaining nodes remain unchanged.

[0059] S204. Input the data corresponding to the nodes of real-time MPEE into the risk assessment model to obtain the risk results; The risk assessment model is trained based on a deep neural network model; the risk results include risk level and risk node.

[0060] In some implementation methods, the risk assessment model construction methods include: Acquire historical risk data for a number of electronic products and production lines; divide training data into task sets; wherein each task set corresponds to a heterogeneous cause-effect graph and risk labeling data for an electronic product production line; A training model is constructed based on a deep neural network, with heterogeneous causal graph nodes as inputs and risk levels and probabilities as outputs. By performing multi-task gradient updates on the model to be trained using training data, a risk assessment model is obtained.

[0061] S205. Based on the risk level and heterogeneous causal graph, root cause location and causal contribution quantification are performed to obtain root cause nodes and their degree of influence; the optimal execution strategy is determined based on the root cause nodes and their degree of influence.

[0062] In some implementations, root cause localization and causal contribution quantification are performed based on risk outcomes and heterogeneous causal graphs, including: Taking the risk node in the risk outcome as the endpoint, and traversing in the opposite direction of the directed causal edge in the heterogeneous causal graph, we obtain several upstream causal paths and upstream MPEE nodes leading to the risk node. Traverse the source nodes in several upstream causal paths; Extract the ATE value from the upstream causal path and determine the number of steps and chain weights in the upstream causal path; calculate the formula... Calculate the influence coefficient (TCC) of all source nodes; where n is the number of causal edges in the l-th indirect causal chain. Let L be the chain weight, and L be the number of indirect causal chains from X to Y. Let be the ATE value of the i-th causal edge in the l-th indirect causal chain; Traverse the source nodes corresponding to the maximum value of the influence level TCC and mark them as root cause nodes.

[0063] It should be pointed out that, The weight is set according to the length of the causal chain. The shorter the length, the greater the weight, so as to avoid the invalid contribution of excessively long causal chains, adapt to the risk propagation characteristics of electronic product manufacturing processes, and when TCC>1, the value is 1.

[0064] In this embodiment, the ATE value of the causal path is extracted, and the path step number n=2 and chain weight ω are determined. l =0.8 (2-step path industrial general weight), substitute into the TCC formula for calculation. Since there is no direct causal relationship between workshop humidity and poor solder joint, TCC=0+(0.92×0.95)×0.8=0.6992, and the root cause node is obtained: workshop humidity, with an influence degree of TCC=0.6992.

[0065] Based on the above technical solution, the electronic product manufacturing risk assessment method based on artificial intelligence provided in this application also includes: determining the optimal execution strategy according to the root cause node and the degree of impact.

[0066] Furthermore, in this embodiment, the method further includes: determining the optimal execution strategy based on the root cause node and the degree of impact.

[0067] In some implementations, the optimal execution strategy is determined based on the root cause node and its degree of impact, including: Extract the manufacturing process knowledge base of electronic products; the manufacturing process knowledge base includes risk suppression strategies for various MPEE node anomalies. Candidate suppression strategies are matched based on root cause nodes; the optimal strategy is selected based on the strategy utility value of each candidate suppression strategy, and the optimal strategy is executed.

[0068] It should be noted that the risk mitigation strategies in the production process knowledge base are classified according to the dimensions of "equipment, process, materials, and environment", and each root cause node corresponds to 1-3 sets of suitable candidate mitigation strategies.

[0069] Furthermore, the strategy utility value can be calculated by weighted summation based on the risk suppression result, intervention cost, and production interruption impact. Among multiple candidate suppression strategies, the strategy with the highest utility value is selected as the optimal strategy. Specifically, the risk suppression result... The risk level can be determined based on the degree of decrease after the strategy is implemented. With the current risk level The higher the ratio, the better the inhibitory effect; intervention cost The actual cost of implementing the execution strategy The maximum acceptable cost of a single intervention by the enterprise The smaller the ratio, the lower the cost; while the impact of generation interruption... Production downtime caused by strategy execution can be assessed. Maximum acceptable interruption time for a single enterprise intervention The smaller the value, the smaller the impact on production.

[0070] In this embodiment, the electronic product manufacturing process knowledge base (categorized by equipment, process, materials, and environment) is retrieved, and two candidate suppression strategies are matched based on root cause nodes and workshop humidity (environmental category): A1: Turn on the workshop dehumidifier and adjust the humidity to the normal process range of 30%RH~70%RH. No need to stop the machine. A2: Turn on the workshop dehumidifier and increase workshop ventilation to quickly reduce humidity to the normal range without stopping the machine.

[0071] The strategy utility value SU is calculated according to the formula. Calculate; where γ+ + =1, The values ​​are set to 0.6~0.7, λ (intervention cost) to 0.2~0.3, and μ (impact of production interruption) to 0.1~0.2. In this embodiment, the weights are set as follows: γ (risk suppression effect) = 0.7, λ (intervention cost) = 0.2, μ (impact of production interruption) = 0.1; the current risk level R0 = 2, the maximum acceptable cost for a single intervention for the enterprise Cmax = 500 yuan, and the maximum acceptable interruption time Tmax = 30 min. The parameter values ​​and SU calculation results for each strategy are as follows: Strategy A1: ΔR=1 (risk decreases from level 2 to level 1), C=20 yuan, T(A1)=0min, SU(A1)=0.7×(1 / 2)-0.2×(20 / 500)-0.1×(0 / 30)=0.342; Strategy A2: ΔR=1 (risk decreases from level 2 to level 1), C=30 yuan, T(A2)=0min, SU(A2)=0.7×(1 / 2)-0.2×(20 / 500)-0.1×(0 / 30)=0.342; Based on the SU value from high to low, strategy A1 was determined to be the optimal execution strategy.

[0072] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, the disclosure, and the appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0073] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for risk assessment in electronic product manufacturing based on artificial intelligence, characterized in that, include: Obtain heterogeneous elements of the electronic product manufacturing process and construct a heterogeneous cause-effect graph and a standard MPEE of electronic product production processes, and construct a standard dataset D0; Real-time MPEE data of the production process is collected using a fixed-time sliding window, and a sliding window dataset D is constructed. t ; Based on the sliding window dataset D t The distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph; The data corresponding to the nodes of real-time MPEE are input into the risk assessment model to obtain the risk result; wherein, the risk assessment model is trained based on a deep neural network model; Root cause localization and causal contribution quantification are performed based on risk level and heterogeneous causal graph to obtain root cause nodes and degree of influence.

2. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 1, characterized in that, The method for constructing the heterogeneous causal graph includes: Extract the MPEE of each process in the production of electronic products; where MPEE includes material node M, process node P, equipment node E, and environment node En. MPEE elements are defined as nodes in a causal graph. Directed causal edges between causal nodes are constructed using a constrained causal model. The weights of the directed causal edges are quantified. The causal graph nodes, directed causal edges, and their corresponding weights are then integrated to obtain a preliminary causal graph. The preliminary causal graph is validated, and the validated preliminary causal graph is marked as a heterogeneous causal graph.

3. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 2, characterized in that, The quantification method for the weights of the directed causal edges includes: Retrieve the cause node X and effect node Y from the cause-effect graph; Determine whether cause node X and effect node Y are consecutive nodes; No, the average causal effect is calculated using the potential outcome framework and labeled as the weight of the directed causal edge; the average causal effect is calculated as follows: ; The average causal effect of node X on node Y; do(X=1) is the action that forces node X to be in an abnormal state; This represents the potential outcome of Y when X=1; Let X be the potential outcome of Y when X=0; do(X=0) is the dynamic that forces node X to the normal state; E[] is the mathematical expectation; Yes, the causal node X is divided into a normal domain X0 and an abnormal domain X1 according to the process threshold; the average causal effect is calculated and marked as the weight of the directed causal edge; wherein, the formula for calculating the average causal effect is... ;in, Let X be the sample size when node X is in the outlier region X1. Let X be the sample size when node X is in the normal domain X0.

4. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 3, characterized in that, The update methods for heterogeneous causal graphs include: The kernel-based MMD algorithm is used to compute the detection of the sliding window dataset D. t The distribution difference between the dataset and the standard dataset D0; the formula for calculating the distribution difference is as follows: ; The sample size of the standard dataset D0, Sliding window dataset D t Sample size; Let i be the MPEE feature vector of the i-th and j-th samples in the standard dataset D0; For the sliding window dataset D t The MPEE feature vectors of the m-th and n-th samples in the dataset; The Gaussian kernel function; Set the distribution offset threshold based on the stability of electronic product manufacturing processes. ; Distribution difference and distribution offset threshold Compare; when the distribution difference is greater than the distribution offset threshold If the distribution differences are analyzed using the feature layer MMD decomposition method, the offset nodes are determined; the heterogeneous causal graph is updated based on the offset nodes; otherwise, the heterogeneous causal graph is saved.

5. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 4, characterized in that, The distribution offset threshold is set according to the stability of electronic product manufacturing process. ,include: Collect historical data on the electronic product manufacturing process where no anomalies occurred, and construct a benchmark dataset D. 0-m ; Using a sliding window on the benchmark dataset D 0-m Self-comparison calculations were performed to obtain the MMD with normal fluctuations. 2 Statistical characteristics; Based on MMD 2 Statistical characteristics are used to calculate the distribution offset threshold; where the calculation formula is: ; For MMD 2 The mean of a statistical characteristic; This refers to the process capability coefficient. This is the quantile correction factor. For MMD 2 Standard deviation of statistical characteristics This represents the sensitivity coefficient to process risks.

6. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 4, characterized in that, The method for determining the offset node includes: Extract the MPEE nodes from the production process to obtain a node list; The proportion of each node's ATE value in the node list is marked as the node weight. A specific offset threshold is calculated based on the node weight and the distribution offset threshold. The formula for calculating the specific offset threshold is: ;in, Let be the node weight of the k-th node; This is the exclusive offset threshold for the k-th node; The distribution offset of each node is calculated by kernelizing the MMD formula; The distribution offset of each node is compared with the exclusive offset threshold, and the nodes whose distribution offset is greater than the exclusive offset threshold are marked as offset nodes.

7. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 6, characterized in that, The root cause localization and causal contribution quantification based on risk outcomes and heterogeneous causal graphs include: Extract risk results; wherein, the risk results include risk level and risk node; Taking the risk node in the risk outcome as the endpoint, and traversing in the opposite direction of the directed causal edge in the heterogeneous causal graph, we obtain several upstream causal paths and upstream MPEE nodes leading to the risk node. Traverse the source nodes in several upstream causal paths; Extract the ATE value from the upstream causal path and determine the number of steps and chain weights in the upstream causal path; calculate the formula... Calculate the influence coefficient (TCC) of all source nodes; where n is the number of causal edges in the l-th indirect causal chain. Let L be the weight of the l-th indirect causal chain, and L be the number of indirect causal chains from X to Y. Let be the ATE value of the i-th causal edge in the l-th indirect causal chain; Traverse the source nodes corresponding to the maximum value of the influence level TCC and mark them as root cause nodes.

8. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 1, characterized in that, Also includes: Determine the optimal execution strategy based on the root cause node and the degree of impact; The process of determining the optimal execution strategy based on the root cause node and the degree of impact includes: Extract the manufacturing process knowledge base of electronic products; the manufacturing process knowledge base includes risk suppression strategies for various MPEE node anomalies. Candidate suppression strategies are matched based on root cause nodes; the optimal strategy is selected based on the strategy utility value of each candidate suppression strategy, and the optimal strategy is executed.

9. The method for risk assessment in electronic product manufacturing based on artificial intelligence according to claim 1, characterized in that, The method for constructing the risk assessment model includes: Acquire historical risk data for a number of electronic products and production lines; divide training data into task sets; wherein each task set corresponds to a heterogeneous cause-effect graph and risk labeling data for an electronic product production line; A training model is constructed based on a deep neural network, with heterogeneous causal graph nodes as inputs and risk levels and probabilities as outputs. By performing multi-task gradient updates on the model to be trained using training data, a risk assessment model is obtained.

10. An artificial intelligence-based electronic product manufacturing risk assessment system, applied to the method described in any one of claims 1-9, characterized in that, The device includes: a preprocessing module and an evaluation processing module; The preprocessing module is used to acquire heterogeneous elements of the electronic product manufacturing process and construct a heterogeneous cause-effect graph and a standard MPEE for electronic product production processes. Real-time MPEE data of the production process is collected using a fixed-time sliding window, and a sliding window dataset D is constructed. t and the standard dataset D0; Based on the sliding window dataset D t The distribution offset between the heterogeneous causal graph and the standard dataset D0 is used to update the heterogeneous causal graph; The assessment and processing module is used to input the data corresponding to the nodes of real-time MPEE into the risk assessment model to obtain the risk result; wherein, the risk assessment model is trained based on a deep neural network model; Root cause localization and causal contribution quantification are performed based on risk level and heterogeneous causal graph to obtain root cause nodes and degree of influence.

Images