A manufacturing-oriented multi-agent collaborative attribution analysis method

Abstract

This invention relates to a multi-agent collaborative attribution analysis method for manufacturing, belonging to the field of artificial intelligence. In this method, when multiple defects are found in a concentrated set of products undergoing quality inspection on a production line, a scheduling agent analyzes the entire set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect. The entire set of process parameters includes the values ​​of various process parameters at different times. Based on the overlap relationship of suspicious process parameters among the suspicious feature subsets corresponding to multiple defects, the scheduling agent generates multiple root cause hypothesis relationships. The scheduling agent triggers multiple analysis agents to analyze the multiple root cause hypothesis relationships respectively, obtaining attribution analysis results. Where all available computing power remains constant, the computing power used when a suspicious process parameter in a root cause hypothesis relationship is analyzed by its corresponding analysis agent is greater than the computing power used when it is analyzed by the scheduling agent.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and relates to a multi-agent collaborative attribution analysis method for the manufacturing industry. Background Technology

[0002] In precision manufacturing, when defects are detected by the quality inspection system on a production line, it is usually necessary to investigate a massive amount of process parameters related to the production line to pinpoint the root cause of the defects. Traditional attribution analysis methods mainly rely on human experience, with engineers manually retrieving historical data from each suspected process and analyzing and attributing the causes through trend chart comparisons, correlation analyses, and other means. This approach is inefficient. Therefore, in highly automated production lines, models (such as intelligent agents) can be used to perform a global analysis of process parameters and defects, thereby improving the efficiency of analysis and attribution.

[0003] However, improving the accuracy of model analysis and attribution remains a research challenge. Summary of the Invention

[0004] In view of this, in order to solve the above problems, the present invention provides a multi-agent collaborative attribution analysis method for the manufacturing industry.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Firstly, a multi-agent collaborative attribution analysis method for manufacturing is provided. This method is applied to a scheduling agent and includes the following steps: When multiple defects are found in a batch of products undergoing quality inspection on a production line, the scheduling agent analyzes the full set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect. The full set of process parameters includes the values ​​of various process parameters at different times. The subset of suspicious features contains suspicious process parameters associated with the corresponding defect, and these suspicious process parameters belong to multiple process parameters. Based on the overlap relationship of suspicious process parameters among the subsets of suspicious features corresponding to multiple defects, the scheduling agent generates multiple root cause hypothesis relationships. Each root cause hypothesis relationship indicates that the corresponding defect may have a root cause relationship with at least one associated suspicious process parameter. The number of suspicious process parameters involved in the multiple root cause hypothesis relationships is less than the full set of process parameters. The scheduling agent triggers multiple analysis agents to analyze the multiple root cause hypothesis relationships respectively, obtaining attribution analysis results. Wherein, with all available computing power remaining constant, the computing power used by a suspicious process parameter in a root cause hypothesis relationship when analyzed by the corresponding analysis agent is greater than the computing power used by the scheduling agent when analyzed.

[0007] Therefore, when facing scenarios with multiple defects occurring concurrently, this solution first analyzes the entire set of process parameters, quickly filtering out suspicious feature subsets associated with each defect, thus achieving initial convergence from the entire set of parameters to suspicious parameters. Then, utilizing the parameter overlap relationships between the suspicious feature subsets of each defect, multiple root cause hypothesis relationships are generated, allowing multiple defects sharing the same suspicious parameter to be aggregated into the same relationship, improving analysis efficiency. Finally, multiple analysis agents are triggered to analyze the root cause hypothesis relationships. With the total available computing power remaining constant, the analysis agents concentrate computing power on a smaller number of suspicious process parameters, resulting in significantly higher computing power per process parameter compared to the full analysis stage. This improves the accuracy of analysis and attribution without increasing hardware resources.

[0008] Optionally, the full set of process parameters includes M types of process parameters, the values ​​of each of the M types of process parameters at N time points within a preset time period, and the time points at which multiple defects occur among the N time points. M and N are both integers greater than 1. The scheduling agent analyzes the full set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect. This includes: the scheduling agent constructing a matrix X and a vector Y based on the full set of process parameters. The N rows of matrix X represent N time points, the M columns of matrix X represent the M types of process parameters, and each element in matrix X represents the value of a corresponding process parameter at a corresponding time point. The vector Y is a 0 / 1 sequence of length N, where each element has a value of 0 or 1 to indicate whether a defect occurs at a corresponding time point among the N time points. The scheduling agent analyzes matrix X and vector Y to obtain a subset of suspicious features corresponding to each defect.

[0009] Therefore, the complete set of process parameters is organized into a matrix X and a vector Y. Rows in matrix X correspond to time points, and columns correspond to process parameters. Vector Y marks the occurrence of defects at each time point. This structure unifies multi-dimensional, time-series process data into a standardized numerical expression, enabling subsequent scanning analysis to be performed on a unified mathematical model. This avoids discrete processing on a parameter-by-parameter and time-point basis, improving processing efficiency and accuracy.

[0010] Optionally, the scheduling agent analyzes matrix X and vector Y to obtain a subset of suspicious features corresponding to each defect, including: the scheduling agent centers each column of matrix X to obtain matrix X', and centers vector Y to obtain vector Y'; the scheduling agent performs a matrix-vector multiplication on matrix X' and vector Y' to obtain a set of suspicious score vectors S; the scheduling agent selects the top K suspicious score vectors from the set of suspicious score vectors S, and the suspicious features corresponding to the top K suspicious score vectors are used to constitute the subset of suspicious features corresponding to each defect.

[0011] Therefore, it can be seen that the centralized processing eliminates the deviation caused by different dimensions and baselines of each process parameter, so that the correlation score reflects the degree of synchronization between the fluctuation of each parameter relative to its own mean and the occurrence of defects, rather than the magnitude of the absolute value. Selecting the top K scores from the score set can quickly lock the suspicious feature subset most likely related to the defects.

[0012] Optionally, matrices X and X' satisfy the following relationship: X' = X - mean(X); where X' is matrix X', X is matrix X, mean(X) is a row vector composed of the mean values ​​of each column of matrix X, and the length of mean(X) is M; vectors Y and Y' satisfy the following relationship: Y' = Y - mean(Y); where Y' is vector Y', Y is vector Y, and mean(Y) is the mean value of all elements in vector Y.

[0013] Therefore, matrix X' is obtained by subtracting the mean of each column of matrix X, and vector Y' is obtained by subtracting the mean of vector Y. The mathematical definition of this centralization operation clarifies the specific methods of matrix and vector preprocessing, so that the values ​​of each process parameter are uniformly converted into deviation values ​​relative to their own historical mean, that is, a unified processing dimension, to improve the accuracy of screening.

[0014] Optionally, the matrix X', vector Y', and the set of suspicious score vectors S satisfy the following relationship: S=(X'ᵀ·Y') / (N-1); where · represents multiplication, / represents division, X'ᵀ is the transpose of matrix X', Y' is vector Y', and S is the set of suspicious score vectors S with a length of M. Each element in the set of suspicious score vectors S represents the correlation score between a corresponding process parameter and the corresponding defect.

[0015] Therefore, the set of suspicious score vectors S is obtained by multiplying the transpose of the centered matrix X' with the centered vector Y' and then dividing by N-1. This formula explicitly defines the calculation method for the correlation score, which is essentially a parallel calculation of the covariance between all process parameters and defect labels. Thus, the normalization operation of dividing by N-1 eliminates the influence of the number of time windows on the absolute value of the score, making the suspicious scores obtained under different defects or different time window lengths comparable.

[0016] Optionally, the scheduling agent generates multiple root cause hypothesis relationships based on the overlapping relationships of suspicious process parameters among suspicious feature subsets corresponding to multiple defects. These relationships include: the scheduling agent associating suspicious process parameters that simultaneously exist in the suspicious feature subsets corresponding to at least two defects with the same root cause hypothesis relationship, wherein at least two defects in the same root cause hypothesis relationship share a root cause hypothesis relationship with the suspicious process parameters; and the scheduling agent associating suspicious process parameters that exist only in the suspicious feature subset corresponding to one defect with a separate root cause hypothesis relationship, wherein a defect in the separate root cause hypothesis relationship has an independent root cause hypothesis relationship with the suspicious process parameters.

[0017] Therefore, the generation of root cause hypotheses is based on the overlap of suspected process parameters among multiple defects. When the same suspected process parameter appears simultaneously in the suspected feature subsets of at least two defects, it is associated with the same root cause hypothesis, indicating that it may be a shared root cause. Suspicious process parameters appearing only in a single defect are associated with a separate hypothesis, indicating that they may be an independent root cause. This binning approach utilizes the correlation information between defects, allowing shared and independent root causes to be processed separately, i.e., targeted coupling and decoupling, to improve the accuracy of subsequent analysis and processing.

[0018] Optionally, for any one of the multiple root cause hypothesis relationships, where each target root cause hypothesis relationship corresponds to one of the multiple analytical agents, the scheduling agent triggers multiple analytical agents to analyze the multiple root cause hypothesis relationships respectively, obtaining attribution analysis results. This includes: the scheduling agent assigning an aggregation level to the target root cause hypothesis relationship based on its complexity, where the aggregation level corresponds to the type of the target analytical agent and also to the number of suspicious process parameters in the target root cause hypothesis relationship analyzed by the target analytical agent in a single analysis; the scheduling agent sending analysis task information to the target analytical agent, which includes the aggregation level and the target root cause hypothesis relationship; and the target analytical agent analyzing the target root cause hypothesis relationship based on a hash function and the aggregation level, obtaining the target attribution analysis results, and sending the target attribution analysis results to the scheduling agent.

[0019] Therefore, the scheduling agent assigns an aggregation level to the target root cause hypothesis based on its complexity. The aggregation level also determines the type of target analysis agent invoked and the number of suspicious process parameters analyzed in a single instance. This allows for matching higher-complexity root cause hypotheses with more capable agent types and larger single-analysis coverage, while lower-complexity buckets are matched with lightweight agents and smaller coverage. Analysis task information is uniformly encapsulated and sent to the target analysis agent, which then performs the analysis based on a hash function and the aggregation level, and returns the results. This mechanism achieves adaptive matching of analysis resources and task complexity, avoiding the problem of insufficient computing power for high-complexity tasks or wasted computing power for low-complexity tasks.

[0020] Optionally, the target analysis agent analyzes the target root cause hypothesis relationship based on the hash function and the aggregation level to obtain the target attribution analysis result, including: the target analysis agent determines the analysis starting offset in the target root cause hypothesis relationship based on the hash function; the target analysis agent takes the analysis starting offset as the analysis starting point and the number of suspicious process parameters corresponding to the aggregation level as the analysis step size, and performs causal analysis on the suspicious process parameters in the target root cause hypothesis relationship in sequence to obtain the target attribution analysis result.

[0021] Therefore, the target analysis agent determines the starting offset within the target root cause hypothesis relationship using a hash function. Then, with the number of single analyses corresponding to the aggregation level as the step size, it sequentially performs causal analysis on the suspicious process parameters within the relationship, starting from this offset. During causal inference, the order of analysis may affect the model's judgment of causal relationships between variables. A fixed starting position and traversal order can easily introduce systematic bias, causing parameters ranked earlier to obtain higher causal confidence. In this scheme, the starting offset is determined by a hash function combined with the current time seed through randomization. This ensures that the starting point and traversal order of the agent's analysis of each suspicious parameter within the root cause hypothesis relationship are different each time the analysis task is executed. This randomization disrupts the fixed analysis order, eliminates the preconceived bias caused by the fixed ranking, and places each suspicious parameter in a fairer order in different triggered analyses, thereby improving the robustness of the attribution conclusions.

[0022] Optionally, the target analysis agent determines the starting offset of the analysis in the target root cause hypothesis relationship based on a hash function, including: the target analysis agent obtaining its own identifier, the relationship identifier of the target root cause hypothesis relationship, and the current time seed; the target analysis agent concatenates its own identifier, the relationship identifier, and the current time seed, calculates the hash function, and takes the modulo to obtain the starting offset of the analysis.

[0023] Therefore, the calculation of the initial offset depends on the target analysis agent's own identifier, the relation identifier of the target root cause hypothesis, and the current time seed. The concatenation of these three identifiers is processed using a hash function, and the modulo operation yields an integer offset within the range of the total number of suspicious process parameters in the bucket. The participation of the self-identifier and relation identifier ensures that different agents receive differentiated offsets when processing different buckets, while the participation of the time seed allows the same agent to obtain different offsets at different times, enhancing the dynamism and flexibility of task allocation. This method eliminates the need for a central node to maintain a global task allocation table, allowing each agent to independently complete its localization.

[0024] The starting offset for analysis satisfies the following relationship: Offset = H(AgentID‖RelationID‖TimeSeed) mod T; where Offset is the starting offset for analysis, H is the hash function, AgentID is its own identifier, RelationID is the relation identifier, TimeSeed is the current time seed, T is the total number of suspicious process parameters in the target root cause hypothesis relation, and ‖ indicates the splicing operation.

[0025] In a second aspect, a control terminal is provided, including a memory and a processor, wherein the memory is used to store computer programs and the processor is used to execute the computer programs to implement the method of the first aspect.

[0026] Thirdly, a multi-agent system is provided, which is configured to perform the method as described in the first aspect.

[0027] The objectives and other advantages of this invention can be realized and obtained through the following description. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0029] Figure 1 This invention provides a schematic diagram of an architecture based on a multi-agent system.

[0030] Figure 2 A flowchart of a multi-agent collaborative attribution analysis method for manufacturing provided by the present invention;

[0031] Figure 3 This is a schematic diagram of the structure of a control terminal provided by the present invention. Detailed Implementation

[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. The accompanying drawings are for illustrative purposes only, representing only schematic diagrams and not actual physical objects, and should not be construed as limiting the present invention. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product form.

[0033] like Figure 1 The diagram illustrates an architecture of a multi-agent system provided in this application. The system includes a scheduling agent and multiple analysis agents.

[0034] The scheduling agent, in the form of a software program, is deployed on a Manufacturing Execution System (MES) server, an Industrial Internet of Things (IIoT) platform server, or a standalone edge computing node. The scheduling agent communicates with the production line data acquisition system via an application programming interface (API) to obtain a complete set of process parameters related to defects. It also interacts with various analysis agents through an internal communication protocol, distributing analysis task information and collecting attribution analysis results. The core functions of the scheduling agent in this embodiment include: performing a first-stage scanning analysis on the complete set of process parameters in a multi-defect concurrent scenario to generate a subset of suspicious features corresponding to each defect; generating multiple root cause hypothesis relationships based on the overlap of suspicious process parameters between the subsets of suspicious features; assigning aggregation levels according to the complexity of each root cause hypothesis relationship, creating corresponding types of analysis agents, and broadcasting root cause hypothesis relationship information and the current time seed to the analysis agents; and summarizing the attribution analysis results returned by each analysis agent and outputting a final attribution report. The scheduling agent itself does not perform deep causal analysis; its computing power consumption is mainly used for matrix operations and task coordination, accounting for a small proportion of the total computing power pool of the system.

[0035] Multiple analytical agents, each an independently running software process or containerized instance, are deployed in computing clusters equipped with graphics processing units (GPUs) or dedicated artificial intelligence (AI) acceleration chips. Analytical agents are categorized into several types based on their aggregation level, each embedding a causal analysis model of varying complexity. Upon creation, each analytical agent independently determines its assigned root cause hypothesis and the initial feature offset within its bucket based on a hash function. It then loads the original sensor data for the corresponding suspected process parameter for causal verification and sends the attribution analysis results back to the scheduling agent. With all available computing power remaining constant, the analytical agents concentrate their computing power on a smaller number of suspected process parameters than the full set, ensuring that the computing power density allocated to a single suspected process parameter is higher than during the scanning phase.

[0036] like Figure 2 As shown in the figure, this application provides a multi-agent collaborative attribution analysis method for the manufacturing industry, which is applied to the system described in the above embodiment.

[0037] The specific process of this method is as follows:

[0038] like Figure 2 As shown, this application provides a multi-agent collaborative attribution analysis method for the manufacturing industry, which is applied to the system described in the above embodiment. The specific process of this method is as follows.

[0039] S201, when multiple defects are found in products being inspected on the production line, the scheduling agent analyzes the full set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect.

[0040] First, the triggering conditions for this step are explained. "Multiple defects appearing in products inspected on the production line" means that during continuous production, the number of defective products detected by the quality inspection process within a preset inspection cycle or within a preset product batch range exceeds a preset quantity threshold, or the proportion of defective products in that batch exceeds a preset proportion threshold. The "multiple defects" refer to at least two defects, which can be the same type of defect appearing in different batches or at different locations, or different types of defects being detected simultaneously or sequentially. For example, two different types of defects, solder joint cracks and shell micropores, are detected simultaneously in the same quality inspection batch; or, for example, solder joint cracks, the same type of defect, are detected in three consecutive batches. The specific value of the preset inspection cycle can be configured according to the production line cycle time and quality inspection frequency. For example, a 3-minute cycle is configured for a fast-cycle surface mount technology production line, and a 15-minute cycle is configured for a slow-cycle final assembly production line. The specific value of the preset quantity threshold can be configured according to the normal quality fluctuation level of the production line. For example, a 2-threshold threshold is configured for a Six Sigma quality level production line, and a 5-threshold threshold is configured for a conventional quality level production line. The specific value of the aforementioned preset ratio threshold can be configured to any value between 1% and 5%, such as 2%. This step is triggered when the number of defect alarms received by the scheduling agent within the preset detection period reaches the preset number threshold, or the defect ratio reaches the preset ratio threshold.

[0041] The complete set of process parameters includes the values ​​of various process parameters at different times. These various process parameters refer to the sum of all monitored process parameters on the production line that may be related to product quality, including but not limited to: temperature parameters (such as reflow soldering zone temperature, mold temperature, injection molding barrel temperature), pressure parameters (such as injection pressure, holding pressure, mold closing pressure), speed parameters (such as pick-and-place machine speed, conveyor belt speed, robotic arm movement speed), vibration parameters (such as pick-and-place machine Z-axis vibration amplitude, spindle vibration frequency), flow parameters (such as cooling water flow rate, raw material feed flow rate), environmental parameters (such as workshop temperature, workshop humidity, cleanliness), electrical parameters (such as power supply voltage, current, power), and equipment status parameters (such as equipment runtime, cumulative runtime after maintenance, number of processing operations after tool replacement). The total number of process parameters in the complete set of process parameters is denoted as M, where M is an integer greater than 1. For example, for a typical surface mount technology production line, M = 3000. The aforementioned different times can be N time points within a preset time period, meaning the full set of process parameters includes the values ​​of each process parameter at N time points within the preset time period. The preset time period refers to the backtracking time window related to the occurrence of defects; that is, the scheduling agent backtracks a certain duration before determining when to trigger this step, and this certain duration is the preset time period. The duration of this preset time period can be determined based on the production cycle and process time of the production line. For example, for a production line with a production cycle of one product per minute, the preset time period can be configured to backtrack 24 hours before the defect detection time, in which case N=1440 and the time granularity is 1 minute; for a production line with a slower production cycle, the preset time period can be configured to backtrack 72 hours, with a time granularity of 5 minutes, in which case N=864.

[0042] The complete set of process parameters also includes the time points at which multiple defects occurred across N time points. For example, each defect has a corresponding defect label. The defect label is used to indicate whether a product manufactured at each of the N time points was subsequently detected by the quality inspection system. For instance, if a product manufactured at a certain time point is subsequently detected for the defect, the defect label for that time point is set to 1; otherwise, it is set to 0. It should be noted that when multiple different types of defects exist (i.e., multiple defects contain different types of defects), a corresponding defect label is constructed for each defect type, with different defect types corresponding to different defect labels.

[0043] The complete set of process parameters is obtained by the scheduling agent from the Manufacturing Execution System (MES) or Industrial Internet of Things (IIoT) platform via an application programming interface (API). The MES collects and stores real-time time-series data of process parameters uploaded by sensors at each workstation, as well as defect detection results from quality inspection workstations. The IIoT platform provides unified equipment data access and storage services, standardizing sensor data of different protocols and formats and storing it in a time-series database. After triggering the attribution analysis process, the scheduling agent requests data from the aforementioned systems via API calls according to preset time periods and defect types to obtain the complete set of process parameters.

[0044] A subset of suspicious features contains suspicious process parameters that are associated with the corresponding defects. These suspicious process parameters belong to multiple process parameters, and the association refers to a statistically significant correlation between the changes in the values ​​of the process parameters and the occurrence of the defects, including but not limited to linear or simple nonlinear correlations. The subset of suspicious features is the result of a preliminary screening of the full set of process parameters, and it contains fewer suspicious process parameters than the full set. For example, if the full set of process parameters contains M=3000 process parameters, after screening, the subset of suspicious features corresponding to each defect contains approximately K suspicious process parameters, where K can range from 5% to 15% of M, for example, K=300.

[0045] The above analysis process will be illustrated with a specific example below.

[0046] Suppose that during the quality inspection process at 14:30 on May 30, 2026, an electronics manufacturing production line detects both solder joint cracks and micropore defects in the casing. The scheduling agent retrieves the full set of process parameters for a preset time period from the manufacturing execution system. This preset time period is 24 hours prior to the defect detection time, with a time granularity of 1 minute (N=1440, M=3000). Taking solder joint cracks as an example, the scheduling agent constructs a matrix X and a vector Y based on the full set of process parameters.

[0047] Matrix X has N rows representing N time points, and M columns representing M process parameters. Each element in matrix X represents the value of a corresponding process parameter at a specific time point. For example, a partial structure of matrix X is shown in Table 1 below:

[0048] Table 1

[0049] 14:31 217.3 0.12 85.2 ... 58.3 14:32 217.5 0.11 85.0 ... 58.1 14:33 222.1 0.15 85.8 ... 58.4 14:34 216.8 0.10 84.9 ... 57.9 ... ... ... ... ... ...

[0050] Vector Y is a 0 / 1 sequence of length N, where each element takes the value 0 or 1, representing whether a weld crack defect occurs at a corresponding time point. For example, a partial structure of vector Y is shown in Table 2 below:

[0051] Table 2

[0052] Time point Are there any cracks in the solder joints? 14:31 0 14:32 0 14:33 1 14:34 0 ... ...

[0053] The scheduling agent analyzes matrix X and vector Y to obtain a subset of suspicious features corresponding to weld crack defects. The specific implementation is as follows.

[0054] The first step is to center each column of matrix X to obtain matrix X', and center vector Y to obtain vector Y'. The purpose of centering is to eliminate the incomparability of various process parameters due to differences in dimensions and baselines, so that the processed data reflects the fluctuation of each parameter relative to its own mean.

[0055] For example, matrices X and X' satisfy the following relationship:

[0056] X' = ​​X - mean(X). Here, X' is matrix X', X is matrix X, and mean(X) is a row vector composed of the mean values ​​of each column of matrix X, with a length of M. Each element in mean(X) is the arithmetic mean of a corresponding process parameter over N time points. For example, if the average temperature of reflow soldering zone 3 over 1440 time points is 217.1℃, then the value of the corresponding column in mean(X) is 217.1. After centering, each element in matrix X' represents the deviation of the corresponding process parameter from its own mean at the corresponding time point. For example, at time point 14:33, the original value of reflow soldering zone 3 is 222.1℃, and the centered value is +5.0℃, indicating that the temperature at that time point is 5.0℃ higher than normal.

[0057] Vectors Y and Y' satisfy the following relationship:

[0058] Y' = Y - mean(Y). Here, Y' is a vector Y', Y is a vector Y, and mean(Y) is the mean of all elements in vector Y, representing the overall occurrence rate of weld joint cracks over N time points. For example, if weld joint cracks occur in 72 out of 1440 time windows, then mean(Y) = 72 / 1440 = 0.05. After centering, positive values ​​in vector Y' correspond to time points where defects occur (value 0.95), while negative values ​​correspond to time points where normal products are produced (value -0.05).

[0059] The second step involves the scheduling agent performing a matrix-vector multiplication on matrix X' and vector Y' to obtain a set of suspicious score vectors S. Matrix X', vector Y', and the set of suspicious score vectors S satisfy the following relationship: S = (X'ᵀ · Y') / (N-1). Here, · represents multiplication, / represents division, X'ᵀ is the transpose of matrix X' with shape M×N, and Y' is vector Y' with shape N×1. Multiplying X'ᵀ and Y' yields a vector of length M, which is then divided by (N-1) to obtain the set of suspicious score vectors S. Dividing by (N-1) is a Bessel correction used to obtain an unbiased estimate of the population covariance in the sample covariance calculation. The set of suspicious score vectors S has a length of M, and each element in S represents the correlation score between a corresponding process parameter and the weld crack defect. The second step is essentially equivalent to simultaneously calculating the sample covariance of all M process parameters and defect labels. For a scale of N=1440 and M=3000, the total computational load is approximately 8.6 × 10⁻⁶. 6 Sub-floating-point operations can be completed in milliseconds on current GPUs. The computational load allocated to each process parameter is approximately 2800 multiply-accumulate operations. This means that at this stage, the computational density allocated to each process parameter is low, sufficient only to discover linear or simple nonlinear statistical correlations between process parameters and defects, but not to determine complex causal relationships.

[0060] Taking a solder joint crack defect as an example, when the defect occurred at 14:33, the centralized value of reflow soldering temperature zone 3 was +5.0℃, and the corresponding value in vector Y' was +0.95. The product of the two contributed a large positive value (4.75). However, when there was no defect at 14:31, the centralized value was +0.2℃, and the corresponding value in vector Y' was -0.05. The product of the two contributed a small negative value (-0.01). After summing the products of all 1440 time points, reflow soldering temperature zone 3 obtained a high suspicion score. Conversely, for parameters unrelated to the defect, the product of the centralized value and Y' was randomly distributed (positive or negative), and the summation approached zero.

[0061] The third step involves the scheduling agent selecting the top K suspicious score vectors from the suspicious score vector set S. The suspicious process parameters corresponding to these top K suspicious score vectors are used to construct the suspicious feature subset corresponding to solder joint crack defects. The value of K can be determined according to a preset ratio (e.g., 10% of M, i.e., K=300) or according to a preset score threshold (e.g., all process parameters with an absolute value greater than 0.3 for suspicious scores). After screening, the suspicious feature subset corresponding to solder joint crack defects includes approximately 300 suspicious process parameters, such as reflow soldering temperature zone 3 (score 82.3), pick-and-place machine Z-axis vibration (score 45.7), mold temperature (score 38.2), and cooling water flow rate (score 31.5). In other words, the suspicious feature subset corresponding to solder joint crack defects contains approximately 300 suspicious process parameters, including reflow soldering temperature zone 3, pick-and-place machine Z-axis vibration, mold temperature, and cooling water flow rate. Similarly, using the above method, it is assumed that a subset of suspicious features corresponding to the micropore defects in the shell can be obtained, such as mold temperature (score 45.1), injection pressure (score 52.3), raw material drying time (score 28.7), etc., about 300 suspicious process parameters. That is, the subset of suspicious features corresponding to the micropore defects in the shell contains about 300 suspicious process parameters such as mold temperature, injection pressure, raw material drying time, etc.

[0062] S202, the scheduling agent generates multiple root cause hypothesis relationships based on the overlapping relationships of suspicious process parameters among suspicious feature subsets corresponding to multiple defects.

[0063] Each root cause hypothesis indicates that the corresponding defect and at least one associated suspected process parameter may have a root cause relationship. A root cause relationship refers to a causal association where the abnormal fluctuation of a suspected process parameter is the fundamental or primary cause that may lead to the corresponding defect. The phrase "may have a root cause relationship" is used here because the subset of suspected features obtained through statistical analysis in S201 only reflects the statistical correlation between process parameters and defects, and cannot yet determine the causal direction. This needs to be finally confirmed by the analytical agent in S203 after in-depth causal verification.

[0064] In this study, the number of suspected process parameters involved in multiple root cause hypothesis relationships is smaller than the total set of process parameters. In other words, after screening in S201 and aggregation in S202, the total number of suspected process parameters entering the subsequent in-depth analysis stage is significantly reduced compared to the size of the total set of process parameters. For example, the total set of process parameters contains 3000 process parameters. After screening in S201, each defect corresponds to approximately 300 suspected process parameters. After deduplication and merging in S202, the total number of non-repeating suspected process parameters involved in all root cause hypothesis relationships is approximately 50 to 100. This gradual convergence of the analysis scope aims to increase the computing power density of individual parameters while maintaining the total computing power.

[0065] It should be noted that S202 is the processing logic in a scenario with multiple concurrent defects. When there is only a single defect, S202 can be simplified to directly associating each suspicious process parameter in the suspicious feature subset corresponding to the defect with an independent root cause hypothesis relationship, that is, each suspicious process parameter constitutes an independent root cause hypothesis relationship with the defect.

[0066] The following section continues the example from S201, providing a detailed explanation of the process for generating root cause hypothesis relationships.

[0067] In the example of S201, the suspicious feature subset corresponding to the solder joint crack defect includes approximately 300 suspicious process parameters such as reflow soldering temperature zone 3 (score 82.3), mold temperature (score 38.2), and cooling water flow rate (score 31.5); the suspicious feature subset corresponding to the shell micropore defect includes approximately 300 suspicious process parameters such as mold temperature (score 45.1), injection pressure (score 52.3), and raw material drying time (score 28.7). The scheduling agent analyzes the overlap between the suspicious feature subsets corresponding to the above two defects, that is, checks whether the same suspicious process parameter appears simultaneously in the suspicious feature subsets of both defects. For example, mold temperature exists in both the suspicious feature subset of the solder joint crack and the suspicious feature subset of the shell micropore, constituting an overlap.

[0068] Based on the aforementioned overlapping relationships, the scheduling agent generates multiple root cause hypothesis relationships, specifically including two cases.

[0069] Scenario 1:

[0070] The scheduling agent associates suspicious process parameters that simultaneously exist in the suspicious feature subsets corresponding to at least two defects with the same root cause hypothesis relationship. At least two defects in this same root cause hypothesis relationship share a root cause hypothesis relationship with the suspicious process parameter. A shared root cause hypothesis relationship means that the anomaly of the suspicious process parameter is assumed to be a common root cause leading to the occurrence of the aforementioned at least two defects. For example, if mold temperature simultaneously exists in the suspicious feature subsets of weld cracks and shell micropores, the scheduling agent generates the first root cause hypothesis relationship. This root cause hypothesis relationship indicates that weld cracks and shell micropores may share a root cause relationship with mold temperature. One possible representation is: Mold temperature (weld crack, shell micropore), meaning that both weld cracks and shell micropores are associated with mold temperature, i.e., weld cracks and shell micropores may share a root cause relationship with mold temperature. The physical meaning of this hypothesis is: abnormal fluctuations in mold temperature may simultaneously cause weld cracks at the welding station and shell micropores at the injection molding station. Similarly, if the cooling water flow rate exists simultaneously in the subset of suspected features of solder joint cracks and another defect (such as a decrease in yield), a corresponding shared root cause hypothesis relationship is generated.

[0071] Scenario 2:

[0072] The scheduling agent associates suspicious process parameters that exist only in a subset of suspicious features corresponding to a single defect with a single root cause hypothesis. Within this single root cause hypothesis, a defect has an independent root cause hypothesis relationship with that suspicious process parameter. An independent root cause hypothesis relationship means that the anomaly of the suspicious process parameter is assumed to be an independent root cause of the defect, with no direct causal relationship to other defects. For example, reflow soldering temperature zone 3 exists only in the subset of suspicious features for solder joint cracks, and not in the subsets of suspicious features for other defects. The scheduling agent generates a separate root cause hypothesis relationship for it, which could be represented as: Reflow Soldering Temperature Zone 3 - Solder Joint Crack, indicating that the solder joint crack and reflow soldering temperature zone 3 may have an independent root cause relationship. Similarly, injection pressure exists only in the subset of suspicious features for micropores in the shell, and a separate independent root cause hypothesis relationship is generated for it.

[0073] After the above processing, the root cause hypothesis relationships generated in this example can be summarized as follows: mold temperature and solder joint cracks and shell micropores constitute a shared root cause hypothesis relationship; reflow soldering temperature zone 3 and solder joint cracks constitute an independent root cause hypothesis relationship; injection pressure and shell micropores constitute an independent root cause hypothesis relationship; and so on, other suspicious process parameters that only exist in a single defect each constitute an independent root cause hypothesis relationship with the corresponding defect.

[0074] Therefore, by generating root cause hypothesis relationships, the scheduling agent decomposes the multi-defect attribution problem into multiple structured subtasks. Each root cause hypothesis relationship corresponds to an independent root cause direction to be verified and only includes suspicious process parameters directly related to that direction. The generation of shared root cause hypothesis relationships utilizes the correlation information between defects. When the same suspicious process parameter appears in the suspicious feature subsets of multiple defects, it is aggregated into a shared hypothesis for analysis, rather than processing the same parameter separately in the independent analysis of each defect. This avoids the repeated investigation of shared parameters and fragmented conclusions that occur when traditional methods analyze shared parameters independently by defect type. At the same time, compared to the full set of process parameters, the total number of suspicious process parameters involved in all root cause hypothesis relationships is significantly reduced.

[0075] S203, the scheduling agent triggers multiple analysis agents to analyze multiple root cause hypothesis relationships respectively, and obtains the attribution analysis results.

[0076] Specifically, assuming all available computing power remains constant, the computing power used by a suspect process parameter in the root cause hypothesis relationship when analyzed by its corresponding analytical agent is greater than the computing power used by the scheduled agent when performing the analysis. In other words, the computing power density allocated to each process parameter is lower in the S201 scanning phase and higher in the S203 analysis phase. The following explanation of the change in computing power density is illustrated with example data.

[0077] Assume the system's total computing power pool contains four A100 GPUs. In the S201 scanning phase, all computing power is used to scan approximately 3000 process parameters, with the computing power density allocated to each parameter recorded as a baseline. In the S203 analysis phase, the same total computing power is focused on approximately 50 to 100 non-repeating suspected process parameters involved in all root cause hypotheses. Because the analysis scope converges from approximately 3000 parameters to approximately 50 to 100 (approximately 1 / 60 to 1 / 30 of the original), under the constraint of unchanged total computing power, the computing power density allocated to each suspected process parameter in the analysis phase is increased by approximately 30 to 60 times compared to the scanning phase. This increase in computing power density allows the analysis agent to load more complex causal analysis models onto each suspected process parameter, use higher-precision raw data, and move from coarse-grained statistical correlations in the scanning phase to fine-grained causal inferences.

[0078] For ease of understanding, for any target root cause hypothesis among multiple root cause hypotheses, the target root cause hypothesis corresponds to one of the multiple analytical agents. It should be noted that the above "one-to-one correspondence" means that one target root cause hypothesis is analyzed by at least one target analytical agent; typically, one target root cause hypothesis corresponds to one target analytical agent. When the complexity of the target root cause hypothesis is high and the number of suspected process parameters involved is large, multiple target analytical agents of the same type can be configured to collaboratively analyze the same target root cause hypothesis. The target analytical agents naturally form a staggered division of labor through the hash mechanism described later.

[0079] The implementation of S203 will be introduced below, taking the example of a target root cause hypothesis relationship corresponding to a target analysis agent.

[0080] The first step is to assign aggregation levels to the target root cause hypothesis relationships based on their complexity.

[0081] The aggregation level has two functions: First, the aggregation level corresponds to the type of the target analytical agent, that is, the aggregation level determines which level of analytical agent is invoked to process the root cause hypothesis; Second, the aggregation level corresponds to the number of suspicious process parameters in the target root cause hypothesis in a single analysis by the target analytical agent, that is, the aggregation level determines the number of parameters covered by the analytical agent in each analysis step when traversing suspicious process parameters.

[0082] In this application embodiment, an exemplary correspondence between the aggregation level and the type of the target analysis agent, the embedded causal analysis model, and the applicable scenario is as follows.

[0083] Aggregation level L1 corresponds to a lightweight analytical agent. The lightweight analytical agent embeds a Granger causality test model. The Granger causality test is a statistical hypothesis testing method based on time series forecasting. Its basic principle is: if adding past values ​​of process parameter X significantly improves the prediction accuracy of defect label Y, and this improvement is statistically significant, then X is considered a Granger cause of Y. This model is suitable for low-complexity scenarios with a single defect and a single suspected parameter. The lightweight analytical agent analyzes only one suspected process parameter per iteration.

[0084] The aggregation level L2 corresponds to the standard analytical agent. The standard analytical agent embeds a Bayesian network structure learning model. A Bayesian network is a directed acyclic graph (DAG) model where nodes represent random variables (including process parameters and defect labels), and directed edges represent probabilistic dependencies between variables. This model is suitable for medium-complexity scenarios with a single defect and a small number of suspicious parameters. The standard analytical agent analyzes two suspicious process parameters per iteration.

[0085] Aggregation level L3 corresponds to an augmentation analysis agent. The augmentation analysis agent embeds transferentropy analysis and a convergent cross mapping (CCM) model. Transferentropy is a nonparametric causal measure based on information theory, independent of predefined functional forms. Convergent cross mapping, based on Takens' embedding theorem, reconstructs the attractor manifold of a system from a single time series, determining causal relationships through cross-prediction capabilities, and exhibits strong detection capabilities for nonlinear and weakly coupled systems. This model is suitable for high-complexity scenarios involving two defects and a moderate number of suspicious parameters. The augmentation analysis agent analyzes three suspicious process parameters per iteration.

[0086] It should be understood that the embodiments of this application are based on three aggregation levels and three corresponding types of analytical agents, but this is not intended to be a limitation. For example, there may be fewer or more aggregation levels.

[0087] The complexity of the target root cause hypothesis relationship is comprehensively evaluated based on one or more of the following factors: the number of defects associated with the target root cause hypothesis relationship (the more defects, the higher the complexity), the number of suspicious process parameters involved in the target root cause hypothesis relationship (the more parameters, the higher the complexity), the severity level of the associated defects (the higher the severity, the higher the complexity), and whether there are cross-process interactions (the existence of cross-process interactions increases the complexity). The scheduling agent can preset complexity scoring rules, calculate the complexity score by weighted summation of the above factors, and then determine the aggregation level of the target root cause hypothesis relationship according to the preset mapping relationship between the scoring interval and the aggregation level.

[0088] Taking the root cause hypothesis relationships generated in example S202, the mold temperature, solder joint crack, and shell micropore form a shared root cause hypothesis relationship, spanning two defects, and the associated defect severity is high. The scheduling agent assesses its complexity as high, assigns it aggregation level L3, and calls the enhanced analysis agent for analysis. The reflow soldering temperature zone 3 and solder joint crack form an independent root cause hypothesis relationship, a single defect and a single suspected parameter. The scheduling agent assesses its complexity as medium, assigns it aggregation level L2, and calls the standard analysis agent for analysis. Similarly, other root cause hypothesis relationships are assigned their respective aggregation levels according to their complexity.

[0089] The second step is to send the analysis task information to the target analysis agent.

[0090] The analysis task information includes the aggregation level and the target root cause hypothesis relationships. The target root cause hypothesis relationships include the relationship identifier, the list of associated defects and defect type identifiers, the list of suspected process parameters involved and the identifiers of each suspected process parameter, and the root cause hypothesis type (shared root cause hypothesis or independent root cause hypothesis). In addition, the analysis task information also includes the current time seed, which is the timestamp of the current moment obtained by the scheduling agent when triggering the analysis task, such as a Unix timestamp, used for subsequent hash calculations.

[0091] The target analysis agent is a standalone software process or containerized instance (such as a Docker container) deployed in a computing cluster with GPUs or dedicated artificial intelligence (AI) acceleration chips. At creation, the target analysis agent has a causal analysis model of the corresponding aggregation level embedded within it. The model's parameters are pre-trained offline using historical production data and then embedded into the agent. During online attribution analysis, the target analysis agent directly loads the pre-trained model for inference, eliminating the need for online training and ensuring analysis efficiency. Once the target analysis agent is created by the scheduled agent and receives the analysis task information, it begins executing the analysis.

[0092] The third step involves the target analysis agent analyzing the root cause hypothesis relationship of the target based on the hash function and the aggregation level, obtaining the target attribution analysis results, and sending the target attribution analysis results to the scheduling agent.

[0093] The target analysis agent first determines the starting offset of the analysis within the target root cause hypothesis relation based on a hash function. The starting offset refers to the initial position number from which the target analysis agent begins causal analysis among the suspect process parameters included in the target root cause hypothesis relation. The suspect process parameters in the target root cause hypothesis relation are arranged in a preset order (e.g., arranged from highest to lowest according to the suspect score calculated in S201, or arranged according to the process parameters' order in the production line). Each suspect process parameter corresponds to a position number, which is consecutively numbered starting from 0. The starting offset indicates that the analysis begins from the suspect process parameter with the offset number. For example, if Offset=0, the analysis starts from the first suspect process parameter; if Offset=3, the analysis starts from the fourth suspect process parameter. Randomizing the starting offset using a hash function, rather than fixedly starting from the first suspect process parameter, disrupts the fixed analysis order, avoiding biases introduced by a rigid analysis order into the causal inference conclusions and improving the robustness of the analysis.

[0094] In one possible implementation, the target analysis agent obtains its own identifier, the relation identifier of the target root cause hypothesis, and the current time seed. The self-identifier is a unique identifier for the target analysis agent, such as a universally unique identifier (UUID) or sequence number assigned by the scheduling agent when creating the target analysis agent. The relation identifier is a unique identifier for the target root cause hypothesis, such as a number generated by the scheduling agent when generating the root cause hypothesis. The target analysis agent concatenates its own identifier, relation identifier, and current time seed into a string, inputs this string into a preset hash function to obtain a hash value, and then takes the modulus of this hash value, where the modulus is the total number T of suspicious process parameters in the target root cause hypothesis, ultimately obtaining the analysis starting offset.

[0095] The initial offset is analyzed to satisfy the following relationship:

[0096] Offset=H(AgentID‖RelationID‖TimeSeed)modT;

[0097] Where Offset is the starting offset of the analysis, which is an integer between 0 and (T-1), indicating which suspicious process parameter in the target root cause hypothesis relationship to start the analysis (the sequence number starts from 0); H is the hash function; AgentID is the self-identifier of the target analysis agent; RelationID is the relation identifier of the target root cause hypothesis relationship; TimeSeed is the current time seed; T is the total number of suspicious process parameters in the target root cause hypothesis relationship; and || represents the string concatenation operation.

[0098] The hash function H can be SHA-256 from the secure hash algorithm (SHA) family, or MD5 or other hash functions with uniform distribution characteristics. Because SHA-256 is sensitive to input and has a uniform output distribution, any tiny difference in AgentID, RelationID, and TimeSeed will cause a significant change in the hash value, thus ensuring the randomness and unpredictability of the starting offset of the analysis.

[0099] The following section will explain the hash calculation and analysis process in detail, taking into account the root cause assumption of mold temperature sharing in example S202.

[0100] Continuing from the example in S202, the shared root cause hypothesis relationship for mold temperature is identified as "R001". The suspected process parameters involved, after screening in S201 and aggregation in S202, specifically include six suspected process parameters: mold temperature, reflow soldering zone 3, pick-and-place machine Z-axis vibration, injection pressure, cooling water flow rate, and raw material drying time, i.e., T=6. It should be noted that although this root cause hypothesis relationship is generated by aggregation using mold temperature as the shared suspected parameter, it also includes independent suspected process parameters that only appear in solder joint cracks and shell micropores, respectively. This is because after the shared root cause hypothesis relationship is generated, other independent suspected process parameters corresponding to the associated defects are also included in the relationship's scope, allowing the analysis agent to comprehensively consider the interaction between shared root causes and independent factors in the same analysis task. The scheduling agent assigns aggregation level L3 to this relationship, and determines an L3-type enhanced analysis agent as the target analysis agent, identified as "L3-Agent-01", with a current time seed, for example, "1717056000".

[0101] The target analysis agent L3-Agent-01 calculates the starting offset for analysis: It concatenates its own identifier "L3-Agent-01", relation identifier "R001", and time seed "1717056000" into the string "L3-Agent-01R0011717056000", inputs it into the SHA-256 hash function to obtain a 256-bit hash value, and takes the modulo 6 of this hash value. Assume the calculated offset is 2. This offset indicates that the analysis begins from the 3rd suspicious process parameter in the target root cause hypothesis relation (numbered from 0, Offset=0 corresponds to the 1st, Offset=1 corresponds to the 2nd, and Offset=2 corresponds to the 3rd).

[0102] It is important to note that this scheme uses hash function randomization to determine the starting offset of the analysis, rather than starting the analysis from the first suspected process parameter. This is significant because it avoids the bias introduced by a fixed analysis order into the causal inference conclusions. During causal inference, when the analysis agent analyzes multiple suspected process parameters, the order in which parameters are analyzed can affect the causal model's judgment of the causal relationship between variables. For example, in some causal discovery algorithms based on attention mechanisms or structure learning, variables that enter the model first may have an advantage in attention weight allocation or graph structure initialization, resulting in higher causal confidence for parameters ranked earlier. Similarly, when there is an interaction effect between multiple suspected parameters, the parameter analyzed first may be identified as the direct cause, while the parameter analyzed later may be identified as an indirect cause or moderating variable, even though the actual causal direction may be exactly the opposite. If each analysis starts from a fixed first parameter and proceeds in a fixed order, the above-mentioned order bias will be systematically reproduced in each analysis, leading to a decrease in the robustness and objectivity of the attribution conclusions.

[0103] This scheme uses a hash function combined with the current time seed to randomly determine the starting offset for analysis. This ensures that even when analyzing the same root cause hypothesis, the starting position of the analysis varies depending on the time seed. The analysis agent begins its analysis from a random starting offset, disrupting the fixed analysis order. This results in each suspicious process parameter being in a different analysis order in different triggers, eliminating systematic bias caused by fixed ordering and thus improving the robustness and objectivity of the attribution conclusions.

[0104] After determining the starting offset, the target analysis agent uses this offset as the starting point and the number of suspicious process parameters corresponding to the aggregation level as the analysis step size. It then sequentially performs causal analysis on the suspicious process parameters in the target root cause hypothesis relationship to obtain the target attribution analysis results. "Sequentially" means following the order of the suspicious process parameters in the target root cause hypothesis relationship, starting with the parameter corresponding to the starting offset and proceeding parameter by parameter. When analyzing the last parameter, if there are still parameters not covered by the remaining step size, the analysis continues from the first parameter, forming a circular traversal, until all suspicious process parameters in the target root cause hypothesis relationship have been analyzed.

[0105] Continuing with the example above, the aggregation level of the target analysis agent L3-Agent-01 is L3, corresponding to an analysis step size of 3 suspected process parameters. Starting from offset = 2 (i.e., the 3rd suspected process parameter), the analysis proceeds sequentially with a step size of 3. Assuming the order of the 6 suspected process parameters in the target root cause hypothesis is: parameter 0 - mold temperature, parameter 1 - reflow soldering temperature zone 3, parameter 2 - pick-and-place machine Z-axis vibration, parameter 3 - injection pressure, parameter 4 - cooling water flow rate, and parameter 5 - raw material drying time, the target analysis agent starts with parameter 2 (pick-and-place machine Z-axis vibration), analyzing parameters 2, 3, and 4 with a step size of 3, completing the first set of analyses; then it continues the circular traversal, analyzing parameters 0 and 1, until all 6 parameters have been analyzed.

[0106] When analyzing each suspicious process parameter, the target analysis agent loads the raw sensor data for that parameter and performs causal verification using the causal analysis model corresponding to its own type. The implementation methods of the causal analysis models for each type of analysis agent are explained in detail below.

[0107] It should be noted that raw sensor data refers to the unprocessed, unaggregated time-series data collected and uploaded by sensors on the production line at millisecond sampling periods. Unlike the minute-level aggregated data used in the S201 scanning phase, raw sensor data retains the true values ​​of process parameters at each sampling moment, enabling it to capture instantaneous abnormal fluctuations, spikes, or transient changes that might be masked by smoothing in aggregated data. These instantaneous anomalies are often the direct cause of defects; therefore, raw sensor data is essential for high-precision causal inference.

[0108] The raw sensor data is acquired by the target analysis agent in the following manner: The target analysis agent retrieves the raw sensor data for the corresponding process parameters from the time-series database of the Manufacturing Execution System (MES) or Industrial Internet of Things (IIoT) platform, based on the identifiers of each suspicious process parameter in the target root cause hypothesis and the preset backtracking time period (e.g., 1 hour, 4 hours, 24 hours prior to the defect detection time). This query can be executed via Structured Query Language (SCL) or API calls, with query conditions including the process parameter identifier, start timestamp, and end timestamp. Each record stored in the time-series database includes: timestamp (accurate to milliseconds), process parameter identifier, parameter value, and sensor identifier. The data sampling frequency depends on the configuration of the corresponding sensor. For example, temperature sensors are typically configured at 10Hz (i.e., sampling once every 100 milliseconds), vibration sensors are typically configured at 100Hz (i.e., sampling once every 10 milliseconds), and pressure sensors are typically configured at 50Hz to 200Hz. Different parameters may have different sampling frequencies. The target analysis agent maintains the original sampling frequency of each parameter when loading data, without resampling or alignment, to preserve the complete time information in the data.

[0109] The following explanation continues from the previous example. For instance, if the target analysis agent needs to analyze the suspicious process parameter of mold temperature, it queries the time series database for the original sensing data of the mold temperature sensor with parameter identifier "mold_temp_sensor_03" within a preset backtracking period. The data returned by the query is in the following format: one record per line, including a timestamp (e.g., "2026-05-30 14:15:00.000", "2026-05-30 14:15:00.010", "2026-05-30 14:15:00.020", etc.), parameter identifier ("mold_temp_sensor_03"), temperature value (e.g., 217.3℃, 217.5℃, 217.6℃, etc.), and sensor identifier ("sensor_mold_temp_03"). For example, if the target analysis agent needs to analyze the suspicious process parameter of Z-axis vibration of the pick-and-place machine, it can query the vibration sensor data with parameter identifier "placer_vibration_z_01". The returned data includes information such as timestamp and vibration amplitude (unit: mm / s).

[0110] The following describes the causal analysis models for various types of analytical agents.

[0111] (a) Lightweight analytical agent (L1 type):

[0112] A lightweight analytical intelligence embeds a Granger causality test model. The Granger causality test is a statistical hypothesis testing method based on time series forecasting, used to determine whether one time series is a Granger cause of another time series. This method is prior art in this application, and its basic principles will not be elaborated upon in the embodiments of this application.

[0113] In this embodiment, the lightweight analytical agent uses the original sensor data of the suspected process parameters and the time series of corresponding defect labels as model input. Regarding the time series of defect labels, as described in S201, the time of defect occurrence is measured in minutes; each minute window corresponds to a 0 or 1 label value to indicate whether the defect occurred in that minute. To correspond with the timestamps of the original sensor data, the analytical agent maps the defect labels from the minute windows to all millisecond-level sampling points within that window. Specifically, if the defect label for a certain minute window is 1, then the defect labels corresponding to all millisecond-level sampling points within that minute window are all set to 1; if it is 0, then they are all set to 0. For example, if the defect label for the minute window 14:33 is 1, then the defect labels for all original sensor data sampling points within that minute (14:33:00.000 to 14:33:59.999) are all 1. After the above mapping, the time series of defect labels and the original sensor data form a one-to-one correspondence in timestamps, serving as the input to the lightweight analytical agent.

[0114] The lightweight analytical agent utilizes a Granger causality test model to examine whether the raw sensor data of a suspected process parameter is a Granger cause of a defect label. The lag order of the model is automatically selected based on the Akaike Information Criterion, and the significance level is preset to 0.05. The model output is the determination of whether the suspected process parameter is a Granger cause of the corresponding defect, and the confidence level determined based on the p-value corresponding to the test statistic. For example, if the p-value is less than 0.01, the confidence level is high, meaning a causal relationship exists between a suspected process parameter and its corresponding defect; if the p-value is between 0.01 and 0.05, the confidence level is medium, meaning a causal relationship between a suspected process parameter and its corresponding defect is possible; if the p-value is greater than 0.05, the parameter cannot be considered a Granger cause, meaning a causal relationship between a suspected process parameter and its corresponding defect does not exist.

[0115] (II) Standard analytical agent (L2 type):

[0116] The standard analytical intelligent body embeds a Bayesian network structure learning model. Bayesian network structure learning is a method for learning probabilistic dependencies between variables from data, which is a prior art in this application. The basic principles of this application will not be elaborated here.

[0117] In this embodiment, the standard analytical agent uses the original sensor data of multiple suspicious process parameters in the target root cause hypothesis relationship and the time series of the corresponding defect labels for each defect as model input. The original sensor data of each suspicious process parameter is retrieved from a time-series database. The time series of each defect label is mapped from a minute window to millisecond-level sampling points and corresponds one-to-one with the timestamp of the original sensor data; details are as described above and will not be repeated here. After discretizing the continuous variables, the standard analytical agent uses a scoring search method to learn the network structure. The scoring function can be the Bayesian Information Criterion (BIC) score, and the search algorithm can be hill climbing. The model output is a directed acyclic graph (DAG). The suspicious process parameter corresponding to the directed edge pointing to the defect label node is identified as the direct cause of the defect. The confidence level of each directed edge is determined based on the conditional probability of the edge or the difference in BIC scores before and after including the edge; the greater the score difference, the higher the confidence level.

[0118] (III) Augmented Analytics Agent (L3 type):

[0119] The enhanced analytical system incorporates transfer entropy analysis and a convergent cross-mapping model. Transfer entropy is a nonparametric causal measurement method based on information theory, and convergent cross-mapping is a causal detection method based on attractor reconstruction in nonlinear dynamic systems. Both methods are existing technologies in this field, and their basic principles will not be elaborated upon in the embodiments of this application.

[0120] In this embodiment, the augmentation analysis agent uses the original sensor data of the suspected process parameters and the time series of the defect labels for each corresponding defect as model input. The original sensor data of the suspected process parameters is obtained from a time-series database. The time series of the defect labels for each defect is mapped from minute windows to millisecond-level sampling points and corresponds one-to-one with the timestamps of the original sensor data; details can be found in the foregoing description and will not be repeated here. The augmentation analysis agent runs transfer entropy analysis and convergent cross-mapping analysis respectively. The model output includes the causal direction and confidence level from the suspected process parameters to their corresponding defects. The confidence level is comprehensively evaluated by the statistical significance of the transfer entropy and the convergence accuracy of the convergent cross-mapping. For transfer entropy, a null hypothesis distribution is generated by randomly shuffling the time series, and the percentile of the actual transfer entropy value in the null hypothesis distribution is calculated. The higher the percentile, the higher the confidence level. For convergent cross-mapping, the convergence value of the prediction accuracy increasing with the sequence length is used as an estimate of the causal strength. The higher the convergence value and the more stable the convergence trend, the higher the confidence level. The final confidence level can be determined by combining the results of the two methods, such as taking the weighted average of the two or the higher one.

[0121] After completing a round of analysis of all suspicious process parameters and corresponding defects in the target root cause hypothesis relationship, the target analysis agent determines the target attribution analysis results based on the confidence level of each suspicious process parameter.

[0122] Specifically, the target analysis agent compares the confidence level of each suspicious process parameter with a preset confidence threshold, and determines whether the suspicious process parameter is the root cause of the corresponding defect based on the comparison result. For example, the preset high confidence threshold is 0.8, and the preset low confidence threshold is 0.5. If the confidence level of a suspicious process parameter for a defect is greater than or equal to 0.8, then the suspicious process parameter is confirmed as the root cause of the defect; if the confidence level is between 0.5 and 0.8, then the suspicious process parameter is marked as a possible root cause of the defect, and further verification is recommended; if the confidence level is less than 0.5, then the suspicious process parameter is excluded as the root cause of the defect.

[0123] The following explanation continues from the previous example. Regarding the shared root cause hypothesis of mold temperature, after analyzing all six suspected process parameters, the L3-type enhanced analysis agent obtains the confidence level of each parameter for solder joint cracks and shell micropores. For example, the confidence level of mold temperature for solder joint cracks is 0.88, and the confidence level for shell micropores is 0.84, both greater than or equal to 0.8, confirming that mold temperature is the common root cause of both defects. The confidence level of reflow soldering temperature zone 3 for solder joint cracks is 0.86, greater than 0.8, confirming that reflow soldering temperature zone 3 is an independent root cause of solder joint cracks; however, the confidence level of reflow soldering temperature zone 3 for shell micropores is 0.32, less than 0.5, ruling out its possibility as a root cause of shell micropores. The confidence level for the Z-axis vibration of the pick-and-place machine on solder joint cracks is 0.62, falling between 0.5 and 0.8, and is marked as a possible root cause, with further verification recommended in subsequent analyses. The confidence level for the micropores in the housing is 0.25, less than 0.5, thus its correlation is excluded. Similarly, the target analysis agent traverses all suspicious process parameters, judging them one by one according to the confidence threshold, and finally outputs the target attribution analysis results for this round of analysis.

[0124] It should be noted that the specific values ​​of the confidence thresholds mentioned above are all illustrative examples. Those skilled in the art can flexibly set them according to the quality control requirements of the actual production line and historical attribution experience. This application does not impose specific limitations on this.

[0125] Furthermore, since random starting offsets in a single analysis may lead to differences in the analysis order, which in turn may affect the confidence level, in order to further improve the stability and reliability of the attribution conclusions, the target analysis agent can repeatedly perform multiple rounds of analysis on the same target root cause hypothesis relationship, using different analysis starting offsets in each round of analysis (such as determining the new analysis starting offset by updating the time seed), and combining the results of multiple analyses to determine the final target attribution analysis result.

[0126] Specifically, when the aggregation level is greater than 1, meaning that more than one suspicious process parameter is analyzed in each input model, after completing the first round of analysis, the target analysis agent updates the current time seed, recalculates the analysis starting offset using the updated time seed, and performs the analysis again based on the new starting offset to obtain the results of the second round of single analysis. The current time seed can be updated by obtaining the current system time as the new time seed, or by incrementing the current time seed used in the previous analysis by a preset value (e.g., by 1). This process is repeated until the number of analyses reaches a preset threshold for the number of repeated analysis rounds. The preset threshold for the number of repeated analysis rounds can be configured according to computing resources and analysis accuracy requirements, for example, 3 rounds, 5 rounds, or 10 rounds.

[0127] The starting offset for each round of analysis varies due to the update of the time seed, allowing each round to traverse different suspected process parameters. For example, the first round of analysis starts at offset 2 (corresponding to the 3rd suspected process parameter), the second round starts at offset 5 (corresponding to the 6th suspected process parameter), and the third round starts at offset 0 (corresponding to the 1st suspected process parameter). By analyzing multiple parameters from different starting points, the order of each traversal is disrupted, and each suspected process parameter is considered comprehensively in different analysis sequences, further eliminating the bias that might be introduced by a single fixed order.

[0128] The target analysis agent synthesizes the results of multiple rounds of single analyses to determine the final target attribution analysis result. The synthesis method includes, but is not limited to, any one or more combinations of the following: First, the confidence levels of the same suspicious process parameter in each round of analysis are averaged, and this average confidence level is used as the final confidence level for that parameter; Second, the number of times the same suspicious process parameter is identified as a root cause in each analysis is counted. If the proportion of times it is identified as a root cause exceeds a preset threshold (e.g., more than half), then the parameter is confirmed as a root cause; Third, the consistency of the causal direction of the same suspicious process parameter in each analysis is tested. If the causal directions are consistent across analyses, the confidence level of the causal direction is increased; if the causal directions are inconsistent, the confidence level is decreased.

[0129] After the above single-round or repeated multiple-round analysis, the target analysis agent obtains the target attribution analysis results and sends them to the scheduling agent. The target attribution analysis results include: relationship identifiers, and whether the root cause hypothesis of each suspected process parameter and the corresponding defect is valid.

[0130] In summary, the scheduling agent first analyzes the entire set of process parameters, quickly filtering out suspicious feature subsets associated with each defect, thus achieving initial convergence from the full set of parameters to the suspicious parameters. Then, utilizing the parameter overlap relationships between the suspicious feature subsets of each defect, multiple root cause hypothesis relationships are generated, allowing multiple defects sharing the same suspicious parameter to be aggregated into the same relationship, improving analysis efficiency. Finally, multiple analysis agents are triggered to analyze the root cause hypothesis relationships. With the total available computing power remaining constant, the analysis agents concentrate computing power on a smaller number of suspicious process parameters, resulting in significantly higher computing power per process parameter compared to the full analysis stage. This improves the accuracy of analysis and attribution without increasing hardware resources.

[0131] Figure 3 This is a schematic diagram of the structure of a control terminal provided in an embodiment of this application. Exemplarily, the control terminal can be a terminal, or a chip (system) or other component or assembly that can be disposed on the terminal. Figure 3 As shown, the control terminal 200 may include a processor 201. Optionally, the control terminal 200 may also include a memory 202 and / or a transceiver 203. The processor 201 is coupled to the memory 202 and the transceiver 203, for example, via a communication bus.

[0132] The following is combined Figure 3 A detailed description of each component of the control terminal 200 is provided below:

[0133] The processor 201 is the control center of the control terminal 200. It can be a single processor or a collective term for multiple processing elements. For example, the processor 201 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement the embodiments of this application, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0134] Optionally, the processor 201 can perform various functions of the control terminal 200 by running or executing software programs stored in the memory 202 and calling data stored in the memory 202, such as performing the aforementioned functions. Figure 2 The method shown.

[0135] In a specific implementation, as one example, the processor 201 may include one or more CPUs, for example... Figure 3 CPU0 and CPU1 are shown in the diagram.

[0136] In a specific implementation, as one example, the control terminal 200 may also include multiple processors, for example... Figure 3 The processor 201 shown is an example. Each of the processors 201 can be a single-core processor or a multi-core processor. Here, "processor" can refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0137] The memory 202 is used to store the software program that executes the solution of this application, and the processor 201 controls its execution. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0138] Optionally, the memory 202 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 202 may be integrated with the processor 201 or exist independently, and may be connected via the interface circuit of the control terminal 200. Figure 3 (Not shown in the image) is coupled to processor 201, but this embodiment does not specifically limit this.

[0139] Transceiver 203 is used for communication with other control terminals. For example, if control terminal 200 is a terminal, transceiver 203 can be used to communicate with a network device or with another terminal device. Alternatively, if control terminal 200 is a network device, transceiver 203 can be used to communicate with a terminal or with another network device.

[0140] Optionally, transceiver 203 may include a receiver and a transmitter. Figure 3 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0141] Optionally, the transceiver 203 can be integrated with the processor 201, or it can exist independently and be controlled via the interface circuit of the terminal 200. Figure 3 (Not shown in the image) is coupled to processor 201, but this embodiment does not specifically limit this.

[0142] Understandable Figure 3 The structure of the control terminal 200 shown does not constitute a limitation on the control terminal. The actual control terminal may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0143] Furthermore, the technical effects of the control terminal 200 can be referred to the technical effects of the methods described in the above method embodiments, and will not be repeated here.

[0144] It should be understood that the processor in the embodiments of this application can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0145] It should also be understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0146] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0147] It should be understood that the term "and / or" in this article 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 existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0148] In this application, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0149] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0150] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0151] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0152] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0153] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0154] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0155] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0156] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-agent collaborative attribution analysis method for manufacturing, characterized in that, The method is applied to a scheduling agent, and the method includes: When multiple defects are found in a group of products being inspected on the production line, the scheduling agent analyzes the full set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect. The full set of process parameters includes the values ​​of various process parameters at different times. The subset of suspicious features contains suspicious process parameters that are associated with the corresponding defect. The suspicious process parameters belong to the various process parameters. The scheduling agent generates multiple root cause hypothesis relationships based on the overlapping relationships of suspicious process parameters among multiple suspicious feature subsets corresponding to the defects. Each root cause hypothesis relationship indicates that the corresponding defect may have a root cause relationship with at least one associated suspicious process parameter. The suspicious process parameters involved in the multiple root cause hypothesis relationships are smaller than the full set of process parameters. The scheduling agent triggers multiple analysis agents to analyze the multiple root cause hypothesis relationships to obtain attribution analysis results; wherein, with all available computing power remaining unchanged, the computing power used by the corresponding analysis agent to perform analysis on a suspicious process parameter in the root cause hypothesis relationship is greater than the computing power used by the scheduling agent to perform analysis.

2. The method according to claim 1, characterized in that, The complete set of process parameters includes M types of process parameters, the values ​​of each of the M types of process parameters at N time points within a preset time period, and the time points at which the multiple defects occur among the N time points, where M and N are both integers greater than 1. The scheduling agent analyzes the complete set of process parameters related to the production line to obtain a subset of suspicious features corresponding to each defect, including: The scheduling agent constructs a matrix X and a vector Y based on the full set of process parameters. The N rows of the matrix X represent the N time points, and the M columns of the matrix X represent the M types of process parameters. Each element in the matrix X represents the value of a corresponding process parameter at a corresponding time point. The vector Y is a 0 / 1 sequence of length N, where each element has a value of 0 or 1 to indicate whether the defect occurs at a corresponding time point among the N time points. The scheduling agent analyzes the matrix X and the vector Y to obtain a subset of suspicious features corresponding to each defect.

3. The method according to claim 2, characterized in that, The scheduling agent analyzes matrix X and vector Y to obtain a subset of suspicious features corresponding to each defect, including: The scheduling agent performs centering processing on each column of the matrix X to obtain matrix X', and performs centering processing on the vector Y to obtain vector Y'; The scheduling agent performs a matrix-vector multiplication on the matrix X' and the vector Y' to obtain a set of suspicious score vectors S; The scheduling agent selects the top K suspicious score vectors from the suspicious score vector set S, and the suspicious features corresponding to the top K suspicious score vectors are used to form a subset of suspicious features corresponding to each defect.

4. The method according to claim 3, characterized in that, The matrix X and the matrix X' satisfy the following relationship: X' = ​​X - mean(X); Where X' is the matrix X', X is the matrix X, mean(X) is the row vector formed by the mean of each column of the matrix X, and the length of mean(X) is M; The vector Y and the vector Y' satisfy the following relationship: Y' = Y - mean(Y); Where Y' is the vector Y', Y is the vector Y, and mean(Y) is the mean of all elements in the vector Y.

5. The method according to claim 3, characterized in that, The matrix X', the vector Y', and the set of suspicious score vectors S satisfy the following relationship: S=(X'ᵀ·Y') / (N-1); Where · represents multiplication, / represents division, X'ᵀ is the transpose of the matrix X', Y' is the vector Y', S is the set of suspicious score vectors S, the length of the set of suspicious score vectors S is M, and each element in the set of suspicious score vectors S represents the correlation score between a corresponding process parameter and the corresponding defect.

6. The method according to claim 1, characterized in that, The scheduling agent generates multiple root cause hypothesis relationships based on the overlap relationship of suspicious process parameters among multiple suspicious feature subsets corresponding to the defects, including: The scheduling agent will associate the suspicious process parameters that exist simultaneously in the suspicious feature subsets corresponding to at least two of the defects with the same root cause hypothesis relationship. At least two of the defects in the same root cause hypothesis relationship share a root cause hypothesis relationship with the suspicious process parameters. The scheduling agent associates the suspicious process parameters that exist only in a subset of suspicious features corresponding to a single defect with a single root cause hypothesis relationship. A single defect in a single root cause hypothesis relationship has an independent root cause hypothesis relationship with the suspicious process parameter.

7. The method according to any one of claims 1-6, characterized in that, For any one of the multiple root cause hypothesis relationships, where the target root cause hypothesis relationship corresponds to one of the multiple analysis agents, the scheduling agent triggers multiple analysis agents to analyze the multiple root cause hypothesis relationships respectively, obtaining attribution analysis results, including: The scheduling agent assigns an aggregation level to the target root cause hypothesis relationship based on the complexity of the target root cause hypothesis relationship. The aggregation level corresponds to the type of the target analysis agent and also corresponds to the number of suspicious process parameters in the target root cause hypothesis relationship analyzed by the target analysis agent in a single analysis. The scheduling agent sends analysis task information to the target analysis agent, the analysis task information including the aggregation level and the target root cause hypothesis relationship; The target analysis agent analyzes the target root cause hypothesis relationship based on the hash function and the aggregation level, obtains the target attribution analysis result, and sends the target attribution analysis result to the scheduling agent.

8. The method according to claim 7, characterized in that, The target analysis agent analyzes the target root cause hypothesis relationship based on the hash function and the aggregation level to obtain the target attribution analysis results, including: The target analysis agent determines the starting offset of the analysis in the target root cause hypothesis relationship based on the hash function; The target analysis agent starts the analysis with the analysis start offset and takes the number of suspicious process parameters corresponding to the aggregation level as the analysis step size. It then performs causal analysis on the suspicious process parameters in the target root cause hypothesis relationship in sequence to obtain the target attribution analysis result.

9. The method according to claim 8, characterized in that, The target analysis agent determines the starting offset for analysis within the target root cause hypothesis relationship based on the hash function, including: The target analysis agent obtains its own identifier, the relation identifier of the target root cause hypothesis relationship, and the current time seed; The target analysis agent concatenates its own identifier, the relationship identifier, and the current time seed, calculates the hash function, and then takes the modulo to obtain the analysis starting offset.

10. The method according to claim 9, characterized in that, The starting offset for the analysis satisfies the following relationship: Offset=H(AgentID‖RelationID‖TimeSeed)modT; Where Offset is the starting offset of the analysis, H is the hash function, AgentID is the self-identifier, RelationID is the relation identifier, TimeSeed is the current time seed, T is the total number of suspicious process parameters in the target root cause hypothesis relation, and || represents the splicing operation.

Images