A method and system for diagnosing the causes of quality problems in a complex product assembly process

Abstract

The application discloses a kind of complex product assembly process quality problem factor diagnosis method and system, it is related to factor diagnosis technical field, when complex product appears quality problem, using reverse bayesian network carries out reverse reasoning, determines the preliminary fault node that causes quality problem, for each preliminary fault node, using trained bayesian network carries out forward reasoning, determines the probability that quality problem is caused by preliminary fault node, according to the probability of each preliminary fault node carries out factor diagnosis, to establish trained bayesian network and reverse bayesian network by way of comprehensive reverse reasoning and forward reasoning, factor diagnosis is carried out to the quality problem in complex product assembly process, and diagnosis precision and efficiency are high, can be well applied to the factor diagnosis of quality problem in complex product assembly process.

Description

Technical Field

[0001] This invention relates to the field of factorial diagnostics, and in particular to a method and system for factorial diagnostics of quality problems in complex product assembly processes based on Bayesian networks. Background Technology

[0002] Root cause analysis of quality problems in the assembly process of complex products is an important means of quality control and assurance in the assembly process of complex products. However, the assembly process of complex products involves numerous quality characteristics, and quality problems arising from various quality characteristics can be transmitted and accumulated along the assembly process. Traditional mechanistic modeling methods have many problems, such as large workload and inaccurate calculation of mechanistic parameters due to unclear mechanisms, making them difficult to apply directly in the assembly process of complex products such as helicopters, airplanes, and rockets.

[0003] Therefore, there is an urgent need for a good quality problem analysis and diagnosis technology that is applicable to complex product assembly processes. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for causal diagnosis of quality problems in complex product assembly processes. By establishing a Bayesian network, the method performs causal diagnosis of quality problems in complex product assembly processes, achieving high diagnostic accuracy and efficiency, and is well-suited for causal diagnosis of quality problems in complex product assembly processes.

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

[0006] A method for causal diagnosis of quality problems in complex product assembly processes, the method comprising:

[0007] When a quality problem occurs in a complex product, a reverse Bayesian network is used for reverse reasoning to determine the initial fault node causing the quality problem; the initial fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed based on a trained Bayesian network;

[0008] For each initial fault node, forward reasoning is performed using the trained Bayesian network to determine the probability that the quality problem is caused by the initial fault node. The method for constructing the trained Bayesian network includes: establishing a directed graph of quality characteristic transmission based on the sequence of each assembly process in the complex product assembly process and the transmission relationship between each quality characteristic of adjacent assembly processes; establishing a Bayesian network based on the directed graph of quality characteristic transmission; using the measurement data of each quality characteristic as input, training the Bayesian network using grey relational analysis and maximum likelihood estimation to obtain a trained Bayesian network; the directed graph of quality characteristic transmission includes a first node and a first edge; the first node is each quality characteristic of each assembly process; the first edge is used to connect two first nodes that have a transmission relationship.

[0009] Cause-based diagnosis is performed based on the probability of each of the initial fault nodes.

[0010] In some embodiments, establishing a Bayesian network based on the directed graph transferred by the quality characteristics specifically includes:

[0011] A Bayesian network is established by using the first node of the directed graph of the quality characteristic transmission as the second node of the Bayesian network, and using the first edge of the directed graph of the quality characteristic transmission as the second edge of the Bayesian network; the Bayesian network includes the second node and the second edge connecting the two second nodes.

[0012] In some embodiments, the step of using measurement data of each quality characteristic as input and training the Bayesian network using grey relational analysis and maximum likelihood estimation to obtain a trained Bayesian network specifically includes:

[0013] Using the measurement data of each quality characteristic as input, the structure of the Bayesian network is trained using grey relational analysis to determine the second side with correlation in the Bayesian network; the second side with correlation is recorded as the third side.

[0014] Using the measurement data of each of the aforementioned quality characteristics as input, the parameters of the Bayesian network are trained using the maximum likelihood estimation method to determine the network parameters of each of the aforementioned third sides, thereby obtaining a trained Bayesian network; the trained Bayesian network includes the second node and the third side connecting the two second nodes.

[0015] In some embodiments, the step of using measurement data of each quality characteristic as input and training the structure of the Bayesian network using grey relational analysis to determine the second edge with correlation in the Bayesian network specifically includes:

[0016] For each of the second sides, the measurement data of the quality characteristics corresponding to the two second nodes connected by the second side are respectively recorded as the first data and the second data. The first data includes a plurality of first measurement values, and the second data includes a plurality of second measurement values.

[0017] Calculate a first deviation value between each of the first measured values ​​and the first first measured value, and calculate a first total deviation value based on all the first deviation values; calculate a second deviation value between each of the second measured values ​​and the first second measured value, and calculate a second total deviation value based on all the second deviation values; calculate the absolute correlation based on the first total deviation value and the second total deviation value;

[0018] Calculate a first rate of change value for each of the first measured values, and calculate a first rate of change deviation value between each of the first rate of change values ​​and the first first rate of change value; calculate a first total rate of change deviation value based on all the first rate of change deviation values; calculate a second rate of change value for each of the second measured values, and calculate a second rate of change deviation value between each of the second rate of change values ​​and the first second rate of change value; calculate a second total rate of change deviation value based on all the second rate of change deviation values; calculate the relative correlation degree based on the first total rate of change deviation value and the second total rate of change deviation value.

[0019] Calculate the comprehensive correlation degree based on the absolute correlation degree and the relative correlation degree; determine whether the comprehensive correlation degree is greater than or equal to a first preset threshold; if yes, the second side has a correlation relationship; if no, the second side does not have a correlation relationship.

[0020] In some embodiments, among the two second nodes connected by the third side, the second node with the assembly process preceding the second node is the cause node, and the second node with the assembly process following the second node is the result node; when the cause node is in the normal range, the result node conforms to a normal distribution; when the cause node is in the abnormal range, the result node conforms to a beta distribution; the network parameters of the third side include the parameters of the normal distribution and the parameters of the beta distribution.

[0021] In some embodiments, the step of using the measurement data of each of the quality characteristics as input to train the parameters of the Bayesian network using the maximum likelihood estimation method to determine the network parameters of each of the third sides specifically includes:

[0022] For each of the third sides, the measurement data of the quality characteristics corresponding to the cause node is recorded as the third data, and the measurement data of the quality characteristics corresponding to the result node is recorded as the fourth data.

[0023] Construct the first maximum likelihood function formula corresponding to the normal distribution; determine the parameters of the normal distribution based on the third data, the fourth data, and the first maximum likelihood function formula;

[0024] Construct the second maximum likelihood function formula corresponding to the beta distribution; determine the parameters of the beta distribution based on the third data, the fourth data, and the second maximum likelihood function formula.

[0025] In some embodiments, the method for constructing the inverse Bayesian network specifically includes:

[0026] Using the second node as a node of the reverse Bayesian network and the third edge as a connecting edge of the reverse Bayesian network, a reverse Bayesian network is obtained; the reverse Bayesian network includes multiple nodes and connecting edges connecting two nodes; among the two nodes connected by the connecting edge, the node with the assembly process preceding the assembly process is the result node, and the node with the assembly process following the assembly process is the cause node.

[0027] In some embodiments, after determining the probability that the quality problem is caused by the initial fault node, the diagnostic method further includes generating a joint probability distribution table based on the probability of each initial fault node.

[0028] In some embodiments, the cause-effect diagnosis based on the probability of each of the initial fault nodes specifically includes:

[0029] For each of the initial fault nodes, if the probability of the initial fault node is lower than the second preset threshold, then the initial fault node is not the cause of the quality problem.

[0030] A cause-and-effect diagnostic system for quality problems in the assembly process of complex products, the diagnostic system comprising:

[0031] The reverse reasoning module is used to perform reverse reasoning using a reverse Bayesian network when a quality problem occurs in a complex product, in order to determine the initial fault node causing the quality problem; the initial fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed based on a trained Bayesian network;

[0032] A forward reasoning module is used to perform forward reasoning using the trained Bayesian network for each initial fault node to determine the probability that the quality problem is caused by the initial fault node. The method for constructing the trained Bayesian network includes: establishing a directed graph of quality characteristic transmission based on the sequence of each assembly process in the complex product assembly process and the transmission relationship between each quality characteristic of adjacent assembly processes; establishing a Bayesian network based on the directed graph of quality characteristic transmission; using the measurement data of each quality characteristic as input, training the Bayesian network using grey relational analysis and maximum likelihood estimation to obtain a trained Bayesian network; the directed graph of quality characteristic transmission includes a first node and a first edge; the first node is each quality characteristic of each assembly process; the first edge is used to connect two first nodes that have a transmission relationship.

[0033] The diagnostic module is used to perform cause-effect diagnosis based on the probability of each of the initial fault nodes.

[0034] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0035] This invention provides a method and system for causal diagnosis of quality problems in the assembly process of complex products. When quality problems occur in complex products, a backward Bayesian network is used for backward reasoning to determine the initial fault nodes causing the quality problems. For each initial fault node, a trained Bayesian network is used for forward reasoning to determine the probability that the initial fault node caused the quality problem. Causal diagnosis is performed based on the probability of each initial fault node. By establishing a trained Bayesian network and a backward Bayesian network, and combining backward reasoning and forward reasoning, causal diagnosis of quality problems in the assembly process of complex products is performed. The diagnosis has high accuracy and efficiency and can be well applied to the causal diagnosis of quality problems in the assembly process of complex products. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the diagnostic method provided in Embodiment 1 of the present invention;

[0038] Figure 2 This is a schematic diagram of the diagnostic method provided in Embodiment 1 of the present invention;

[0039] Figure 3This is a schematic diagram of the directed graph for mass characteristic transfer provided in Embodiment 1 of the present invention;

[0040] Figure 4 This is a schematic diagram of the reverse quality characteristic transfer directed graph provided in Embodiment 1 of the present invention;

[0041] Figure 5 This is a schematic diagram of the mass characteristics of the wheel hub provided in Embodiment 1 of the present invention;

[0042] Figure 6 This is a system block diagram of the diagnostic system provided in Embodiment 2 of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] The purpose of this invention is to provide a method and system for causal diagnosis of quality problems in complex product assembly processes. By establishing a Bayesian network, the method performs causal diagnosis of quality problems in complex product assembly processes, achieving high diagnostic accuracy and efficiency, and is well-suited for causal diagnosis of quality problems in complex product assembly processes.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1:

[0047] Traditional mechanism-based modeling methods suffer from numerous problems, such as high workload and inaccurate calculation of mechanism parameters due to unclear mechanisms. They are difficult to apply directly in the assembly process of complex products such as helicopters, airplanes, and rockets. Therefore, this embodiment establishes a Bayesian network to perform cause-effect diagnosis of quality problems in the assembly process of complex products. Compared with mechanism-based modeling methods, this method requires less workload and has higher diagnostic accuracy and efficiency.

[0048] This embodiment provides a method for diagnosing quality problems in the assembly process of complex products, such as... Figure 1 and Figure 2 As shown, the diagnostic method includes:

[0049] S1: When a quality problem occurs in a complex product, a reverse Bayesian network is used for reverse reasoning to determine the initial fault node causing the quality problem; the initial fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed based on a trained Bayesian network.

[0050] S2: For each initial fault node, forward reasoning is performed using the trained Bayesian network to determine the probability that the quality problem is caused by the initial fault node; the method for constructing the trained Bayesian network includes: establishing a directed graph of quality characteristic transmission based on the sequence of each assembly process in the complex product assembly process and the transmission relationship between each quality characteristic of adjacent assembly processes; establishing a Bayesian network based on the directed graph of quality characteristic transmission; using the measurement data of each quality characteristic as input, training the Bayesian network using grey relational analysis and maximum likelihood estimation to obtain a trained Bayesian network; the directed graph of quality characteristic transmission includes a first node and a first edge; the first node is each quality characteristic of each assembly process; the first edge is used to connect two first nodes that have a transmission relationship.

[0051] S3: Perform cause-effect diagnosis based on the probability of each of the initial fault nodes.

[0052] Here, this embodiment provides a detailed description of the construction process of the trained Bayesian network:

[0053] (1) Based on the sequence of each assembly process in the assembly process of complex products and the transmission relationship between each quality characteristic of adjacent assembly processes, a directed graph of quality characteristic transmission is established.

[0054] Complex product assembly processes are multi-level assembly processes, consisting of multiple assembly steps arranged sequentially. Each assembly step involves multiple quality characteristics. Assembly errors are introduced from each quality characteristic of each assembly step and propagate along the assembly process. Therefore, when constructing a directed graph of quality characteristic propagation in complex product assembly processes, the method used in this embodiment is as follows: Each quality characteristic of each assembly step is taken as a first node. Based on the sequence of each assembly step in the complex product assembly process and the propagation relationship between the quality characteristics of adjacent assembly steps, two first nodes with a propagation relationship in adjacent assembly steps are determined. A first edge is drawn between the two first nodes with a propagation relationship, with the direction of the first edge pointing from the first node of the assembly step preceding the first node of the assembly step following the first node, thus establishing a directed graph of quality characteristic propagation. If the propagation relationship cannot be accurately determined, each first node preceding the assembly step and each first node following the assembly step can be connected by the first edge.

[0055] Based on the above method, this embodiment can express the complex product assembly process as follows: Figure 3 The directed graph showing the transfer of quality characteristics is shown. Figure 3 Medium, W.S. k For assembly process k, QC j Let j represent quality characteristic. Figure 3 This indicates that the final assembly quality of a complex product is formed by the progressive transmission of quality characteristics from each assembly step along the assembly process. Based on this, this embodiment can transform the factorial diagnosis of quality problems in the assembly process of complex products into a conditional probability solution process, that is:

[0056] p(WS n QC j Unqualified | WS k QC p (Failure) indicates the final assembly process WS n QC in j The quality problem was caused by WS k QC in p The probability of it occurring.

[0057] (2) Establish a Bayesian network based on the directed graph of quality characteristics.

[0058] A Bayesian network is constructed by using the first node of the directed graph with quality property propagation as the second node of the Bayesian network, and the first edge of the directed graph with quality property propagation as the second edge of the Bayesian network. That is, the Bayesian network is constructed by... Figure 3 Each first node in the network is transformed into a second node in the Bayesian network. Figure 3 Each first edge in the network is transformed into a second edge of the Bayesian network, thus constructing a complex product assembly process as a Bayesian network composed of quality characteristics transmitted along the assembly process. The Bayesian network includes second nodes and second edges connecting two second nodes. The direction of the second edge is from the second node with the assembly process preceding it to the second node with the assembly process following it.

[0059] (3) Using the measurement data of each quality characteristic as input, the Bayesian network is trained by the grey relational analysis method and the maximum likelihood estimation method to obtain the trained Bayesian network.

[0060] The prerequisite for using Bayesian networks for probabilistic solutions is to train the Bayesian network using sample data, including structure learning and parameter learning, to determine whether there are correlations between quality characteristics and the degree of these correlations. Considering that complex product assembly processes are typically small-batch manufacturing scenarios, the resulting sample data is limited and insufficient to meet the sample training requirements of Bayesian networks. To address the problems of poor model learning performance and long learning cycles caused by insufficient sample quantity, this embodiment introduces grey relational analysis into the structure learning process of Bayesian networks.

[0061] (3.1) Using the measurement data of each quality characteristic as input, the structure of the Bayesian network is trained by the grey relational analysis method to determine the second side with the correlation in the Bayesian network, and the second side with the correlation is recorded as the third side.

[0062] Grey relational analysis is a method that uses measurement data of quality characteristics to determine the relationships between them, as detailed below:

[0063] 1) For each second side, the measurement data of the quality characteristics corresponding to the two second nodes connected by the second side are recorded as the first data and the second data, respectively. The first data includes multiple first measurement values, and the second data includes multiple second measurement values.

[0064] The number of first measurements included in the first data and the number of second measurements included in the second data can be the same or different.

[0065] The measurement data can be represented as follows:

[0066] X i =[x i (1), ...,x i (n)] T ;

[0067] Among them, X i x represents a vector consisting of all measured values ​​of the i-th quality characteristic, i.e., the measurement data of the i-th quality characteristic; i (n) represents the nth measurement value of the i-th quality characteristic.

[0068] Assuming there are a total of m+1 quality characteristics in all assembly processes, the measurement data of all quality characteristics can be expressed as:

[0069]

[0070] Matrix M is called the original measurement matrix OMM.

[0071] 2) Calculate the first deviation value between each first measurement value and the first first measurement value, and calculate the first total deviation value based on all first deviation values; calculate the second deviation value between each second measurement value and the first second measurement value, and calculate the second total deviation value based on all second deviation values; calculate the absolute correlation based on the first total deviation value and the second total deviation value.

[0072] The method for calculating the deviation value is as follows:

[0073] For the measurement data X of the i-th quality characteristic i =[x i (1),...,x i (n)] TCalculate the k-th measurement value x respectively. i (k) (k ranges from 1 to n) and the first measured value x i (1) Deviation value

[0074]

[0075] All deviations included in all quality characteristics can be used to form a zero-return matrix M. 0 ,as follows:

[0076]

[0077] The total deviation of the i-th quality characteristic |s i The formula for calculating | is as follows:

[0078]

[0079] Based on this, the first total deviation value and the second total deviation value can be calculated in this embodiment.

[0080] Absolute correlation ε 0i The calculation formula is:

[0081]

[0082] Where |s0| is the first total deviation value; |s i | represents the second total deviation value.

[0083] 3) Calculate the first rate of change value for each first measured value, and calculate the first rate of change deviation value between each first rate of change value and the first first rate of change value. Calculate the first total rate of change deviation value based on all first rate of change deviation values. Calculate the second rate of change value for each second measured value, and calculate the second rate of change deviation value between each second rate of change value and the first second rate of change value. Calculate the second total rate of change deviation value based on all second rate of change deviation values. Calculate the relative correlation degree based on the first total rate of change deviation value and the second total rate of change deviation value.

[0084] In this embodiment, the proportion value d of the i-th quality characteristic can be set. i It is the reciprocal of the first measured value of the i-th quality characteristic:

[0085]

[0086] Multiplying the k-th measurement value by this ratio gives the k-th rate of change value. Therefore, the k-th rate of change value x′ for the i-th quality characteristic is... i The formula for calculating (k) is:

[0087] x′ i (k)=xi (k)*d i .

[0088] The original rate of change matrix M′ can then be constructed as follows:

[0089]

[0090] Deviation value of the k-th rate of change of the i-th quality characteristic The calculation formula is:

[0091]

[0092] Where, x′ i (1) is the first rate of change value of the i-th quality characteristic.

[0093] Then the zero-return matrix M′ can be constructed. 0 ,as follows:

[0094]

[0095] The total rate of change deviation of the i-th quality characteristic |s′ i The formula for calculating | is as follows:

[0096]

[0097] Based on this, the first total rate of change deviation value and the second total rate of change deviation value can be calculated in this embodiment.

[0098] Relative correlation R 0i The calculation formula is:

[0099]

[0100] Where |s′0| is the deviation value of the first total rate of change; |s′ i | represents the deviation value of the second total rate of change.

[0101] 4) Calculate the comprehensive correlation degree based on the absolute correlation degree and the relative correlation degree; determine whether the comprehensive correlation degree is greater than or equal to the first preset threshold; if yes, the second side has a correlation relationship; if no, the second side does not have a correlation relationship.

[0102] The formula for calculating the overall correlation degree ψ is:

[0103] ψ=αε 0i +(1-α)R 0i ;

[0104] Here, α is an empirical value. If it is 0.5, then the absolute correlation and the relative correlation each account for half of the influence. A first preset threshold is set according to the actual situation and experience. If the overall correlation is above the first preset threshold, then the two quality characteristics connected by the second side are considered to be correlated; otherwise, the two quality characteristics connected by the second side are considered not to be correlated.

[0105] Based on the above process, the structure learning of the Bayesian network can be completed, clarifying which quality characteristics are related, and thus determining the second side with the relationship.

[0106] (3.2) Using the measurement data of each quality characteristic as input, the parameters of the Bayesian network are trained using the maximum likelihood estimation method to determine the network parameters of each third side, and the trained Bayesian network is obtained. The trained Bayesian network includes the second node and the third side connecting the two second nodes.

[0107] In this embodiment, the parameter learning process of the Bayesian network is completed using the maximum likelihood estimation method. Of the two second nodes connected by the third edge, the second node with the assembly process preceding the third node is the cause node, and the second node with the assembly process following the third node is the result node. For the cause-result node group constructed in the structure learning, the data distribution pattern of the result node is considered in two cases: when the cause node is in the normal range, the result node conforms to a normal distribution Z ~ N(μ, σ). 2 When the cause node is in the abnormal range, the result node conforms to the beta distribution Z ~ B(α, α), then the network parameters of the third side include the parameters of the normal distribution and the parameters of the beta distribution.

[0108] Then, using the measurement data of each quality characteristic as input, the parameters of the Bayesian network are trained using the maximum likelihood estimation method. The network parameters for each third side can include:

[0109] 1) For each third edge, the measurement data of the quality characteristic corresponding to the cause node is recorded as the third data, and the measurement data of the quality characteristic corresponding to the result node is recorded as the fourth data.

[0110] 2) Construct the formula for the first maximum likelihood function corresponding to the normal distribution, and determine the parameters of the normal distribution based on the third data, the fourth data, and the formula for the first maximum likelihood function.

[0111] The formula for the first maximum likelihood function is as follows:

[0112]

[0113]

[0114]

[0115]

[0116] Among them, z lower and z upper These represent the upper and lower deviations of the random variable z, respectively.

[0117] There are already many software programs that run the maximum likelihood estimation method. After determining the formula for the first maximum likelihood function, the parameters of the normal distribution can be determined by inputting the third data, the fourth data, and the formula for the first maximum likelihood function into the software.

[0118] 3) Construct the formula for the second maximum likelihood function corresponding to the beta distribution, and determine the parameters of the beta distribution based on the third data, the fourth data, and the formula for the second maximum likelihood function.

[0119] The formula for the second maximum likelihood function is as follows:

[0120]

[0121] Where x is a random variable.

[0122] There are already many software programs that run the maximum likelihood estimation method. After determining the formula for the second maximum likelihood function, the third and fourth data and the second maximum likelihood function formula can be input into the software to determine the parameters of the beta distribution.

[0123] Based on the above method, the parameters (μ, σ) of the actual distribution of the result nodes under normal / abnormal conditions corresponding to the cause node of each third edge can be estimated. 2 By using (α, β) and (α, β), the parameters of the Bayesian network can be learned, and the degree of correlation between various quality characteristics can be clarified.

[0124] This embodiment can obtain a well-trained Bayesian network that expresses the mutual influence of quality characteristics along the assembly process of a complex product, and transforms the influence and transmission between quality characteristics into probabilities.

[0125] The factor analysis process for quality problems in complex product assembly can be viewed as the process of discovering quality problems in the final stage of product assembly and tracing back to potentially related quality characteristics, such as... Figure 4 This is a reverse directed graph of quality characteristic transmission. Therefore, in this embodiment, based on the construction logic of Bayesian networks, the reverse probability distribution is calculated to generate a reverse Bayesian network. Querying the reverse Bayesian network can obtain the possible failure rate of each cause node when the state of a certain result node (i.e., the first node of the last assembly process) or result node set (i.e., the set consisting of all the first nodes of the last assembly process) is abnormal.

[0126] Specifically, in this embodiment, the method for constructing the inverse Bayesian network may include: using a second node as a node in the inverse Bayesian network, and using a third edge as a connecting edge in the inverse Bayesian network to obtain the inverse Bayesian network. The inverse Bayesian network includes multiple nodes and connecting edges connecting two nodes. Among the two nodes connected by the connecting edge, the node with the assembly process preceding the assembly process is the result node, and the node with the assembly process following the assembly process is the cause node. That is, compared to the trained Bayesian network, the inverse Bayesian network has the same structure and network parameters as the trained Bayesian network, only the direction of the edges changes. In the trained Bayesian network, the third edge points from the second node with the assembly process preceding the second node with the assembly process following the second node, while the connecting edge of the inverse Bayesian network points from the node with the assembly process following the second node with the assembly process preceding the second node.

[0127] This embodiment utilizes the existing law of total probability and the fundamental theorem of conditional probability to deduce the conditional probability of the inverse Bayesian network as follows:

[0128]

[0129] in, This represents the probability that quality characteristic X1 is qualified. Represents quality characteristic X n The probability of passing; This indicates that when quality characteristic X i When qualified, quality characteristic X i-1 The probability of being qualified; the value of the above probability can be directly calculated by Bayesian network software, that is, it becomes the conditional probability of inferring the cause of the quality characteristics of the upstream process from the result of the quality characteristics of the downstream process.

[0130] After constructing the trained Bayesian network and inverse Bayesian network, the factorial diagnostic process in this embodiment may include:

[0131] (1) When a quality problem occurs, reverse reasoning is performed based on the reverse Bayesian network to determine the initial fault node that caused the quality problem.

[0132] This embodiment determines the quality characteristics of the failure in the last assembly process based on the quality problems that occur. Then, it uses an inverse Bayesian network to perform inference calculations to determine the preliminary fault node. The preliminary fault node can be the first node of any assembly process other than the last assembly process. The assembly process of the preliminary fault node is determined by the user. For example, the first node of an important assembly process can be selected as the preliminary fault node according to actual needs, or the first node of the first assembly process can be selected as the preliminary fault node.

[0133] (2) For each initial fault node, use the trained Bayesian network to perform forward reasoning to determine the probability that the quality problem is caused by the initial fault node.

[0134] After reverse reasoning, this embodiment then performs forward reasoning using a trained Bayesian network, comparing the results with the reverse reasoning results to achieve forward reasoning confirmation. The steps for forward reasoning confirmation are as follows: An initial fault node is arbitrarily selected, and its failure rate is set to 100%. The failure rate of each resulting node is calculated using the Bayesian formula of the trained Bayesian network. Each resulting node is updated. If a resulting node is not the node of the last assembly process, it is changed to a cause node, and reasoning continues to the next layer of result nodes until the probability of a quality problem caused by the initial fault node is deduced. This process is repeated until all initial fault nodes are traversed, and the probability of all initial fault nodes is obtained.

[0135] (3) Perform cause analysis based on the probability of each initial fault node.

[0136] For each initial fault node, if the probability of the initial fault node is lower than the second preset threshold, then the initial fault node is not the cause of the quality problem; if the probability of the initial fault node is higher than or equal to the second preset threshold, then the initial fault node is the cause of the quality problem; the initial fault node with the highest probability is the main cause of the quality problem.

[0137] After determining the probability that a quality problem is caused by an initial fault node, the diagnostic method in this embodiment further includes generating a joint probability distribution table based on the probability of each initial fault node. This embodiment ultimately outputs the joint probability distribution table, the trained Bayesian network, and the inverse Bayesian network, thereby obtaining a highly reliable factorial analysis result for the quality problem.

[0138] This embodiment combines grey relational analysis with Bayesian network structure learning, mapping the quality characteristics of complex product assembly processes to nodes and edges of a Bayesian network. Grey relational analysis achieves efficient structure learning with a small sample size, and combines backward inference with forward inference in the process of factor analysis of quality problems, effectively improving the reliability and efficiency of factor analysis of quality problems.

[0139] The complex products in this embodiment can be helicopters, hubs, airplanes, and rockets.

[0140] Example 1: Factor Analysis of Quality Problems in Helicopter Structural Assembly Process

[0141] (1) Decomposition of helicopter structural assembly process and extraction of quality characteristics of each process;

[0142] The helicopter structural assembly process can be divided into three steps: component assembly, part assembly, and fuselage internal assembly.

[0143] 1) Component assembly

[0144] Component assembly specifically includes:

[0145] The front fuselage frame beam assembly has five quality characteristics: three sets of mating hole coaxiality, assembly surface step difference, and rivet position accuracy.

[0146] The assembly of the mid-fuselage platform has eight quality characteristics: coaxiality of four sets of mating holes, step difference of three assembly surfaces, and position of rivet holes.

[0147] The transition platform assembly has 3 assembly surface steps, 2 sets of mating holes, and 5 quality characteristics related to coaxiality.

[0148] This process has a total of 18 quality characteristics, from Qc1 to Qc. 18 .

[0149] 2) Component assembly

[0150] Component assembly specifically includes:

[0151] The lower assembly of the front fuselage has 6 sets of assembly intersection hole position accuracy, with a total of 6 quality characteristics;

[0152] The mid-fuselage assembly has 6 sets of assembly intersection hole position accuracy and 1 hole symmetry, totaling 7 quality characteristics;

[0153] The transition section assembly has one set of intersection hole position accuracy requirements, totaling one quality characteristic.

[0154] This process has a total of 14 quality characteristics, Qc 19 To Qc 32 .

[0155] 3) Internal assembly of the fuselage

[0156] The internal assembly of the fuselage specifically includes: one set of fairing intersection hole position accuracy, two sets of movable door fixed end connection hole position accuracy, and two sets of tail hatch connection hole position accuracy, totaling five quality characteristics, Qc. 33 To Qc 37 .

[0157] (2) Using the above 37 quality characteristics as nodes, construct a preliminary Bayesian network;

[0158] (3) Using the historical measurement data of the above 37 quality characteristics, begin the structure learning and parameter learning of the Bayesian network;

[0159] (4) After completing the learning, adjust the position of the two sets of connecting holes at the fixed end of the movable door to Qc. 34 With Qc 35 Nodes are removed from the Bayesian network because they are not affected by the quality characteristics of the preceding process, resulting in a trained Bayesian network consisting of 35 quality characteristics.

[0160] (5) Construct a reverse Bayesian network;

[0161] (6) When the next assembly process detects any of the above 35 quality characteristics, especially from Qc 19 To Qc 37 (excluding Qc) 34 With Qc 35 When quality problems occur in the subsequent 17 quality characteristics, bidirectional inference is performed using the constructed trained Bayesian network and inverse Bayesian network to finally identify the cause node, that is, the upstream quality characteristic node that causes quality problems in the downstream of the assembly process.

[0162] This example shows that when learning network structure, it may be found that some downstream quality characteristics are unrelated to upstream quality characteristics, that is, the problem with the quality characteristic is caused only by factors related to the process itself.

[0163] Example 2: Root cause analysis of quality problems in the wheel assembly process

[0164] (1) Extraction of quality characteristics;

[0165] The number of quality characteristics in the wheel hub assembly process is relatively small, consisting of seven, such as... Figure 5 As shown, ① represents the X-direction deviation of the first positioning fixture, ② represents the Y-direction deviation of the first positioning fixture, ③ represents the X-direction deviation of the second positioning fixture, ④ represents the Y-direction deviation of the second positioning fixture, and ⑤ represents the radius.

[0166] ⑥ represents the area, and ⑦ represents the flatness. These seven quality characteristics are named Qc1 to Qc7.

[0167] (2) Construct a Bayesian network using the above 7 quality characteristics as nodes;

[0168] (3) Using the historical measurement data of the above 7 quality characteristics, begin the structure learning and parameter learning of the Bayesian network;

[0169] In this example, since there is no concept of multi-level processes, the structure learning process calculates the correlation between any two of the seven quality characteristics. After structure learning, it is found that Qc7 has no direct relationship with the other six quality characteristics, so it is removed. Then, maximum likelihood estimation and historical measurement data are used to learn the parameters of the Bayesian network, resulting in a trained Bayesian network.

[0170] (4) Construct a reverse Bayesian network;

[0171] In this example, both the trained Bayesian network and the inverse Bayesian network have 6 nodes, but each node has a large number of connection edges, indicating that the connections between the various quality characteristics in this example are very close. This is also related to the obvious physical contact between the various elements in the assembly process, which is in line with expectations.

[0172] (5) After the trained Bayesian network and its inverse Bayesian network are completed, bidirectional inference can be performed to find the corresponding cause quality characteristics when assembly quality problems occur again.

[0173] This embodiment uses the assembly process of complex products, such as helicopters, as a case study, and proposes a causal diagnosis method for quality problems in complex product assembly processes. This method integrates grey relational analysis into the training process of Bayesian networks. Under conditions of small batches of samples and poor sample data quality, it automatically calculates the correlations between various quality characteristics based on quality characteristic measurement data, thereby generating a well-trained Bayesian network for the complex product assembly process. Based on this, a rapid causal tracing method for quality problems based on inverse Bayesian networks is proposed, improving the ability to trace the root causes of quality problems in multi-stage assembly conditions of small-batch complex products.

[0174] Compared with the prior art, the diagnostic method of this embodiment has the following beneficial effects:

[0175] (1) Effectively solved the problem of constructing a factorial model for quality problems in small batches of complex product assembly.

[0176] Grey relational modeling technology was adopted, which can determine the correlation between quality characteristics in complex product assembly process under small batch sample conditions. This solves the problem that the previous Bayesian network construction process required a large amount of data, making it difficult to apply to the factorial analysis of quality problems in complex product assembly process.

[0177] (2) It effectively solves the unified quantification of cause-effect diagnosis for quality problems in complex product assembly processes and has high reliability.

[0178] A Bayesian network for the transmission of quality characteristics in complex product assembly processes and a corresponding inverse Bayesian network for the cause analysis of quality problems were constructed. By combining inverse and forward inference, a joint probability distribution table of quality problems can be calculated. By querying this joint probability distribution table, the causes of quality problems can be quickly located, guiding quality control personnel to handle them quickly. This not only improves the efficiency of quality problem cause analysis, but also enhances the reliability of quality problem cause analysis through the combination of bidirectional inference.

[0179] Example 2:

[0180] This embodiment provides a cause-and-effect diagnostic system for quality problems in complex product assembly processes, such as... Figure 6As shown, the diagnostic system includes:

[0181] The reverse reasoning module M1 is used to perform reverse reasoning using a reverse Bayesian network when a quality problem occurs in a complex product, in order to determine the initial fault node causing the quality problem; the initial fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed based on a trained Bayesian network;

[0182] The forward reasoning module M2 is used to perform forward reasoning using the trained Bayesian network for each initial fault node to determine the probability that the quality problem is caused by the initial fault node. The method for constructing the trained Bayesian network includes: establishing a directed graph of quality characteristic transmission based on the sequence of each assembly process in the complex product assembly process and the transmission relationship between each quality characteristic of adjacent assembly processes; establishing a Bayesian network based on the directed graph of quality characteristic transmission; using the measurement data of each quality characteristic as input, training the Bayesian network using grey relational analysis and maximum likelihood estimation to obtain a trained Bayesian network; the directed graph of quality characteristic transmission includes a first node and a first edge; the first node is each quality characteristic of each assembly process; the first edge is used to connect two first nodes that have a transmission relationship.

[0183] The diagnostic module M3 is used to perform cause-effect diagnosis based on the probability of each of the initial fault nodes.

[0184] The same or similar parts between the various embodiments in this specification can be referred to mutually. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0185] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method of diagnosing a factor of a quality problem in a complex product assembly process, characterized by, The diagnostic method comprises: When a quality problem occurs in a complex product, reverse reasoning is performed by using a reverse Bayesian network to determine a preliminary fault node causing the quality problem; the preliminary fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed according to a trained Bayesian network; For each preliminary fault node, forward reasoning is performed by using the trained Bayesian network to determine a probability that the quality problem is caused by the preliminary fault node; the construction method of the trained Bayesian network comprises: a quality characteristic transmission directed graph is established according to an order of each assembly process in a complex product assembly process and a transmission relationship between quality characteristics of adjacent assembly processes; a Bayesian network is established according to the quality characteristic transmission directed graph; the trained Bayesian network is obtained by using grey correlation analysis and maximum likelihood estimation to train the Bayesian network with measurement data of each quality characteristic as input; the quality characteristic transmission directed graph comprises first nodes and first edges; the first nodes are each quality characteristic of each assembly process; the first edges are used to connect two first nodes having a transmission relationship; Factorial diagnosis is performed according to the probability of each preliminary fault node; The Bayesian network is established according to the first nodes of the quality characteristic transmission directed graph as second nodes of the Bayesian network and the first edges of the quality characteristic transmission directed graph as second edges of the Bayesian network; the Bayesian network comprises the second nodes and second edges connecting two second nodes; The trained Bayesian network is obtained by using grey correlation analysis to train the structure of the Bayesian network with measurement data of each quality characteristic as input and determining second edges having a correlation relationship in the Bayesian network; the second edges having the correlation relationship are recorded as third edges; Network parameters of each third edge are determined by using maximum likelihood estimation to train parameters of the Bayesian network with measurement data of each quality characteristic as input, and the trained Bayesian network is obtained; the trained Bayesian network comprises the second nodes and third edges connecting two second nodes. The trained Bayesian network is obtained by using grey correlation analysis to train the structure of the Bayesian network with measurement data of each quality characteristic as input and determining second edges having a correlation relationship in the Bayesian network; the second edges having the correlation relationship are recorded as third edges; For each second edge, measurement data of quality characteristics corresponding to two second nodes connected by the second edge are recorded as first data and second data respectively; the first data comprises a plurality of first measurement values, and the second data comprises a plurality of second measurement values; 2. The diagnostic method according to claim 1, characterized in that, ​ ​ calculating a first deviation value of each of the first measurement values from a first one of the first measurement values and a first total deviation value from all of the first deviation values; calculating a second deviation value of each of the second measurement values from a first one of the second measurement values and a second total deviation value from all of the second deviation values; and calculating an absolute correlation degree from the first total deviation value and the second total deviation value; calculating a first rate of change value of each of the first measurement values and a first rate of change deviation value of each of the first rate of change values from a first one of the first rate of change values and a first total rate of change deviation value from all of the first rate of change deviation values; calculating a second rate of change value of each of the second measurement values and a second rate of change deviation value of each of the second rate of change values from a first one of the second rate of change values and a second total rate of change deviation value from all of the second rate of change deviation values; and calculating a relative correlation degree from the first total rate of change deviation value and the second total rate of change deviation value; calculating a comprehensive correlation degree from the absolute correlation degree and the relative correlation degree; and determining whether the comprehensive correlation degree is greater than or equal to a first preset threshold value; if yes, the second edge has a correlation relationship; and if no, the second edge does not have a correlation relationship.

3. The diagnostic method according to claim 1, characterized in that, of the two second nodes connected by the third edge, the second node before the assembly procedure is a cause node, and the second node after the assembly procedure is a result node; when the cause node is in a normal interval, the result node conforms to a normal distribution; when the cause node is in an abnormal interval, the result node conforms to a beta distribution; and the network parameters of the third edge include parameters of the normal distribution and parameters of the beta distribution.

4. The diagnostic method according to claim 3, characterized in that, The training of the parameters of the Bayesian network by using the maximum likelihood estimation method with the measurement data of each quality characteristic as input includes the following steps: For each third edge, the measurement data of the quality characteristic corresponding to the cause node is recorded as third data, and the measurement data of the quality characteristic corresponding to the result node is recorded as fourth data. A first maximum likelihood function formula corresponding to the normal distribution is constructed; and the parameters of the normal distribution are determined according to the third data, the fourth data and the first maximum likelihood function formula. A second maximum likelihood function formula corresponding to the beta distribution is constructed; and the parameters of the beta distribution are determined according to the third data, the fourth data and the second maximum likelihood function formula.

5. The diagnostic method according to claim 1, characterized in that, The construction method of the reverse Bayesian network includes the following steps: The second nodes are taken as nodes of the reverse Bayesian network, and the third edges are taken as connecting edges of the reverse Bayesian network to obtain a reverse Bayesian network; the reverse Bayesian network includes a plurality of nodes and connecting edges connecting two nodes; of the two nodes connected by the connecting edge, the node before the assembly procedure is a result node, and the node after the assembly procedure is a cause node.

6. The diagnostic method according to claim 1, characterized in that, After determining the probability of the quality problem being caused by the preliminary fault node, the diagnostic method further comprises: generating a joint probability distribution table according to the probability of each preliminary fault node.

7. The diagnostic method according to claim 1, characterized in that, The factorial diagnosis according to the probability of each preliminary fault node specifically comprises: For each preliminary fault node, if the probability of the preliminary fault node is lower than a second preset threshold, the preliminary fault node is not the cause of the quality problem.

8. A complex product assembly process quality problem causal diagnosis system, characterized in that, The diagnostic system comprises: a reverse reasoning module, configured to, when a quality problem occurs in a complex product, perform reverse reasoning by using a reverse Bayesian network to determine a preliminary fault node causing the quality problem; the preliminary fault node is a node of the reverse Bayesian network; the reverse Bayesian network is constructed according to a trained Bayesian network; a forward reasoning module, configured to, for each preliminary fault node, perform forward reasoning by using the trained Bayesian network to determine the probability of the quality problem being caused by the preliminary fault node; the construction method of the trained Bayesian network comprises: establishing a quality characteristic transmission directed graph according to the order of each assembly process in the assembly process of the complex product and the transmission relationship between each quality characteristic of adjacent assembly processes; establishing a Bayesian network according to the quality characteristic transmission directed graph; taking the measurement data of each quality characteristic as input, training the Bayesian network by using a grey correlation analysis method and a maximum likelihood estimation method to obtain a trained Bayesian network; the quality characteristic transmission directed graph comprises a first node and a first edge; the first node is each quality characteristic of each assembly process; the first edge is used to connect two first nodes having a transmission relationship; a diagnostic module, configured to perform factorial diagnosis according to the probability of each preliminary fault node; wherein the establishment of the Bayesian network according to the quality characteristic transmission directed graph specifically comprises: taking the first node of the quality characteristic transmission directed graph as a second node of the Bayesian network, and taking the first edge of the quality characteristic transmission directed graph as a second edge of the Bayesian network, to establish a Bayesian network; the Bayesian network comprises the second node and a second edge connecting two second nodes; wherein the training of the Bayesian network by taking the measurement data of each quality characteristic as input, and by using the grey correlation analysis method and the maximum likelihood estimation method, to obtain a trained Bayesian network specifically comprises: taking the measurement data of each quality characteristic as input, training the structure of the Bayesian network by using the grey correlation analysis method to determine the second edge having a correlation relationship in the Bayesian network; the second edge having the correlation relationship is recorded as a third edge; taking the measurement data of each quality characteristic as input, training the parameters of the Bayesian network by using the maximum likelihood estimation method to determine the network parameters of each third edge, to obtain a trained Bayesian network; the trained Bayesian network comprises the second node and the third edge connecting two second nodes.

Images