Complex product assembly quality interpretable prediction method considering model uncertainty
Through Bayesian neural network and SHAP interpretability analysis, the accuracy and interpretability problems of assembly quality prediction in complex manufacturing systems are solved, and accurate prediction and uncertainty analysis of complex product assembly quality are achieved.
Patent Information
- Application Number
- CN202510877059.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-27
- Publication Date
- 2025-10-10
AI Technical Summary
Existing assembly quality prediction methods for complex manufacturing systems suffer from insufficient prediction accuracy and poor interpretability when faced with multimodal data distribution and dynamic changes, making it difficult to achieve precise control and quantifiable quality correlation analysis.
An uncertainty-aware Bayesian neural network prediction model is used, combined with SHAP interpretability, to establish a mapping relationship between process parameters and final quality indicators in the assembly process. Through uncertainty quantification modules and sensitivity analysis, interpretable prediction of the assembly quality of complex products is achieved.
It achieves accurate prediction of the assembly quality of complex products, provides model uncertainty distribution analysis and contribution evaluation of assembly process parameters, and improves the interpretability and reliability of predictions.
Smart Images
Figure CN120764359A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of complex product assembly process quality control, and particularly relates to a complex product assembly quality interpretable prediction method considering model uncertainty. BACKGROUND
[0002] The manufacturing of high-end mechanical products, including diesel engines, CNC machine tools and large aircraft, constitutes a complex engineering system characterized by dynamic process changes, limited data availability and multi-modal assembly characteristics. In addition, assembly operations account for 60-70% of the workload throughout the product life cycle, covering the design, manufacturing and testing stages, so its optimization is crucial to production efficiency and quality assurance. For example, on a diesel engine assembly line, process quality control is gradually compromised due to equipment performance drift, human intervention and environmental fluctuations. These factors trigger the accumulation of deviations in successive assembly stages, which can initially be ignored upstream changes escalate through an enhanced propagation mechanism. The resulting amplification cascade, which is captured through fragmented parameters, disrupts the integrity of the data and propagates uncertainty to the consistency of the final product quality. Therefore, assembly quality-oriented prediction with enhanced interpretability is crucial to establishing quantifiable process quality correlations, enabling online parameter optimization while considering the randomness of complex assembly processes.
[0003] Currently, existing research has demonstrated the potential of machine learning and deep learning methods to establish quantitative process quality correlations in complex manufacturing systems. When faced with the multi-modal data distribution characteristics of the manufacturing environment, these machine learning methods exhibit inherent limitations in terms of accuracy. Existing deep learning methods perform well in terms of prediction accuracy, but this improvement in performance often comes at the expense of visibility and interpretability. Therefore, current methods have limitations in terms of prediction accuracy, model interpretability and their practical applicability, and have few successful implementations in real-world applications.
[0004] Currently, the following difficulties still exist in the assembly quality-oriented prediction modeling of complex manufacturing systems:
[0005] 1) Assembly error transferability: Assembly errors at each stage of the assembly process are accumulated through the assembly line to the final product quality, while existing prediction modeling methods often assume fixed data distribution when dealing with various dynamic working conditions. The uncertainty in the model architecture is difficult to quantify, resulting in a decline in prediction reliability.
[0006] 2) Unpredictable Final Quality: The assembly environment is dynamic and volatile, with numerous uncertainties. Quality issues are often not discovered until the testing phase, leaving the assembly process uncontrolled and unresolved, leading to varying final quality within the same batch of products. Existing methods achieve accurate predictions for manufacturing processes such as machining, assembly, and testing, but these quality prediction methods face challenges in balancing performance and interpretability, making it difficult to achieve precise quality control of complex assemblies.
[0007] Therefore, it is necessary to study an interpretable prediction method for complex product assembly quality that considers model uncertainty to address the shortcomings of existing technologies and to solve or alleviate one or more of the above problems. Summary of the Invention
[0008] (1) Technical issues to be resolved
[0009] In response to existing technical problems, the present invention provides an interpretable prediction method for complex product assembly quality taking into account model uncertainty. Historical data of the diesel engine assembly process is collected to establish an uncertainty-aware Bayesian neural network prediction model. The prediction model is used to characterize the mapping relationship between process parameters in the assembly process and the final quality indicators, and a Bayesian neural network uncertainty quantification module is designed. The dynamic decision-making modeling method for assembly quality is then applied to complex manufacturing systems to output the predicted value of the final quality indicator, the uncertainty distribution in the deep learning algorithm, and the contribution value of each assembly process parameter.
[0010] (2) Technical solution
[0011] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:
[0012] A method for interpretable prediction of complex product assembly quality considering model uncertainty includes the following steps:
[0013] S1. Collect historical data of complex product assembly process;
[0014] S2, establishes an uncertainty-aware Bayesian neural network prediction module based on historical data;
[0015] S3. Use a prediction model to characterize the mapping relationship between process parameters in the assembly process and the final quality indicators, and establish a Bayesian neural network uncertainty quantification module;
[0016] S4, using SHAP interpretability to relate uncertainty bounds to feature contributions;
[0017] S5. Based on the uncertainty-aware Bayesian neural network prediction module and the Bayesian neural network uncertainty quantification module, a global sensitivity analysis and a local interpretability module are constructed. The sensitivity analysis outputs the confidence of the predicted value, and the local interpretability outputs the contribution mapping relationship.
[0018] Preferably, the S1 further includes:
[0019] Collect historical data of the diesel engine assembly process, perform data preprocessing on the historical data, and perform statistical processing on typical manufacturing process parameters;
[0020] Among them, statistical processing includes calculating the mean, standard deviation and range of data characteristics.
[0021] Preferably, the specific steps of establishing the uncertainty-aware Bayesian neural network prediction model in S2 include:
[0022] S2.1. Establish an uncertainty-aware Bayesian neural network prediction model. First, define a prior distribution p(ω i |β)~N(0,1), where ω i represents the weight of each process parameter, β represents the initial step set of process parameters, and then the posterior distribution of the model output is defined as Represents the parameter distribution of the network output, and the input assembly process parameters are represented as x i , the predicted final quality is
[0023] S2.2, based on the prior distribution defined in S2.1 and the posterior distribution of the model output, using the principle of variational inference, the variational posterior probability distribution p(ω i ) and the posterior distribution of the model output KL divergence of the difference value The final quality output of the Bayesian neural network is where f(x i ,ω i ) is the prediction function of the model.
[0024] Preferably, the S2 further includes:
[0025] S2.3, based on the final quality output of the Bayesian neural network in S2.2 and the results of multiple forward propagations, the posterior distribution of the model output and the prior distribution p(ω i The final quality and confidence interval of the prediction model are calculated using the expected, variance, and deviation values.
[0026] Preferably, the S2 further includes:
[0027] S2.4. Based on the predicted value and confidence interval of the final quality in S2.3, the final quality and assembly process parameters are correlated to achieve accurate prediction of the final assembly quality.
[0028] Preferably, the S3 further includes:
[0029] S3.1. Build a deep probabilistic architecture of an uncertainty-aware Bayesian neural network to output the final quality prediction value and the uncertainty ratio in the architecture.
[0030] S3.2. Establish a Bayesian neural network uncertainty quantification module and add residual perturbation ε~N(0,δ 2 I), quantifying uncertainty in model architecture and data.
[0031] Preferably, the S3 further includes:
[0032] S3.2.1. Develop quantitative methods corresponding to the uncertainty of the excess distribution;
[0033] Defined as Where c is the total number of assembly features, m is the number of test samples, and y is the final quality value of the actual test;
[0034] S3.2.2. Establish a quantitative method for epistemic uncertainty;
[0035] Defined as The posterior discreteness of the parameters reflects the uncertainty of the quantitative model about the true value of the process parameters;
[0036] S3.2.3. Establish a method to quantify arbitrary uncertainty;
[0037] Defined as Where ε represents the residual variable, Represents the model output value of the residual perturbation.
[0038] Preferably, the S4 further includes: S4.1, defining the contribution of each assembly quality parameter in, Defined as the global interaction score between assembly quality parameters and process characteristics, f(R∪{x i}) As the contribution of assembly quality features to global prediction, N represents the set of assembly quality parameters, R is the subset of the feature set, and ! represents the factorial;
[0039] S4.2. Define the local feature contribution correlation as Quantify the contribution of local feature interactions to prediction uncertainty and establish an optimal relationship between feature importance and uncertainty boundary.
[0040] Preferably, the method further comprises:
[0041] Establish an uncertainty quantification feedback mechanism, classify different uncertainties according to their sources, and dynamically adjust parameter weights based on the uncertainty distribution of assembly quality traceability;
[0042] Apply the explainable prediction method of assembly quality to complex manufacturing systems, establish quantifiable process-quality relationships, and achieve accurate prediction of final quality indicators.
[0043] (3) Beneficial effects
[0044] The beneficial effects of the present invention are as follows: the present invention provides an interpretable prediction method for complex product assembly quality taking into account model uncertainty, which establishes an uncertainty-aware Bayesian neural network prediction model by collecting historical data of the diesel engine assembly process; and uses the prediction model to characterize the mapping relationship between process parameters in the assembly process and the final quality indicators, thereby providing a final quality indicator prediction method based on a deep learning algorithm, model uncertainty distribution analysis, and assembly process parameter contribution evaluation for the quality of complex products. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 A schematic flow chart of a method for interpretable prediction of complex product assembly quality considering model uncertainty provided by the present invention;
[0046] Figure 2 An uncertainty quantification flow chart of an interpretable prediction method for complex product assembly quality considering model uncertainty provided by the present invention;
[0047] Figure 3 An uncertainty distribution diagram in assembly quality prediction modeling of an interpretable prediction method for complex product assembly quality considering model uncertainty provided by the present invention;
[0048] Figure 4 An assembly feature contribution graph of a diesel engine in an embodiment of an interpretable prediction method for complex product assembly quality considering model uncertainty provided by the present invention;
[0049] Figure 5 A comparison chart of the final quality prediction results of a diesel engine in an embodiment of an explainable prediction method for complex product assembly quality considering model uncertainty provided by the present invention. DETAILED DESCRIPTION
[0050] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.
[0051] Example 1
[0052] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides an interpretable prediction method for complex product assembly quality considering model uncertainty, including the following steps:
[0053] S1. Collect historical data of complex product assembly process;
[0054] S2, establishes an uncertainty-aware Bayesian neural network prediction module based on historical data;
[0055] S3. Use a prediction model to characterize the mapping relationship between process parameters in the assembly process and the final quality indicators, and establish a Bayesian neural network uncertainty quantification module;
[0056] S4, using SHAP interpretability to relate uncertainty bounds to feature contributions;
[0057] S5. Based on the uncertainty-aware Bayesian neural network prediction module and the Bayesian neural network uncertainty quantification module, a global sensitivity analysis and a local interpretability module are constructed. The sensitivity analysis outputs the confidence of the predicted value, and the local interpretability outputs the contribution mapping relationship.
[0058] In this embodiment, S1 further includes:
[0059] Collect historical data of the diesel engine assembly process, perform data preprocessing on the historical data, and perform statistical processing on typical manufacturing process parameters;
[0060] Among them, statistical processing includes calculating the mean, standard deviation and range of data characteristics.
[0061] The specific steps of establishing the uncertainty-aware Bayesian neural network prediction model in S2 of this embodiment include:
[0062] S2.1. Establish an uncertainty-aware Bayesian neural network prediction model. First, define a prior distribution p(ω i |β)~N(0,1), where ω i represents the weight of each process parameter, β represents the initial step set of process parameters, and then the posterior distribution of the model output is defined as Represents the parameter distribution of the network output, and the input assembly process parameters are represented as x i , the predicted final quality is
[0063] S2.2, based on the prior distribution defined in S2.1 and the posterior distribution of the model output, using the principle of variational inference, the variational posterior probability distribution p(ω i ) and the posterior distribution of the model output KL divergence of the difference value The final quality output of the Bayesian neural network is where f(x i ,ω i ) is the prediction function of the model.
[0064] In this embodiment, S2 further includes:
[0065] S2.3, based on the final quality output of the Bayesian neural network in S2.2 and the results of multiple forward propagations, the posterior distribution of the model output and the prior distribution p(ω i The final quality and confidence interval of the prediction model are calculated using the expected, variance, and deviation values.
[0066] Preferably, the S2 further includes:
[0067] S2.4. Based on the predicted value and confidence interval of the final quality in S2.3, the final quality and assembly process parameters are correlated to achieve accurate prediction of the final assembly quality.
[0068] In this embodiment, S3 further includes:
[0069] S3.1. Build a deep probabilistic architecture of an uncertainty-aware Bayesian neural network to output the final quality prediction value and the uncertainty ratio in the architecture.
[0070] S3.2. Establish a Bayesian neural network uncertainty quantification module and add residual perturbation ε~N(0,δ 2 I), quantifying uncertainty in model architecture and data.
[0071] In this embodiment, S3 further includes:
[0072] S3.2.1. Develop quantitative methods corresponding to the uncertainty of the excess distribution;
[0073] Defined as Where c is the total number of assembly features, m is the number of test samples, and y is the final quality value of the actual test;
[0074] S3.2.2: Develop methods for quantifying epistemic uncertainty;
[0075] Defined as The posterior discreteness of the parameters reflects the uncertainty of the quantitative model about the true value of the process parameters;
[0076] S3.2.3. Establish a method to quantify arbitrary uncertainty;
[0077] Defined as Where ε represents the residual variable, Represents the model output value of the residual perturbation.
[0078] In this embodiment, S4 further includes: S4.1. Defining the contribution of each assembly quality parameter in, Defined as the global interaction score between assembly quality parameters and process characteristics, f(R∪{x i}) As the contribution of assembly quality features to global prediction, N represents the set of assembly quality parameters, R is the subset of the feature set, and ! represents the factorial;
[0079] S4.2. Define the local feature contribution correlation as Quantify the contribution of local feature interactions to prediction uncertainty and establish an optimal relationship between feature importance and uncertainty boundary.
[0080] The method described in this embodiment also includes:
[0081] Establish an uncertainty quantification feedback mechanism, classify different uncertainties according to their sources, and dynamically adjust parameter weights based on the uncertainty distribution of assembly quality traceability;
[0082] Apply the explainable prediction method of assembly quality to complex manufacturing systems, establish quantifiable process-quality relationships, and achieve accurate prediction of final quality indicators.
[0083] Example 2
[0084] In this embodiment, the present invention is described in detail below with reference to a dataset of a diesel engine assembly process collected in practice, which mainly includes more than 100 assembly process parameters on a diesel engine assembly line, 172 process quality characteristics including crankshaft rotation torque, axial clearance, piston protrusion height, etc., and 24 final performance indicators including power, torque, exhaust temperature, exhaust pressure, etc., through specific implementation methods. Figure 1 shown.
[0085] The preprocessing steps for the actual collected diesel engine assembly process data are as follows:
[0086] Statistical processing of typical production process parameters is required, including calculation of key statistical data such as mean, standard deviation, range, etc.
[0087] Through the data cleaning process, abnormal data generated during data collection and transmission are systematically filtered. Abnormal data is directly deleted.
[0088] For a small number of missing values, mean imputation is performed to facilitate subsequent modeling and analysis.
[0089] Furthermore, the specific steps for establishing an interpretable prediction method for complex product assembly quality with model uncertainty are as follows:
[0090] (1) A prediction model is used to characterize the mapping relationship between process parameters and final quality indicators in the assembly process, and a Bayesian neural network uncertainty quantification module is designed;
[0091] (2) Define a Bayesian neural network with uncertainty perception, set the hyperparameter optimization range, learning rate and KL weight, and introduce three types of uncertainty in the assembly quality prediction modeling process: aleatoric uncertainty (inherent noise in the assembly process dataset), epistemic uncertainty (uncertainty in the neural network model structure and parameters), and out-of-distribution uncertainty (uncertainty when the model faces unknown data distribution);
[0092] (3) Combine the assembly quality prediction method with SHAP interpretability to achieve the integration of interpretable prediction methods for assembly quality of complex products with model uncertainty.
[0093] The specific steps to establish an uncertainty-aware Bayesian neural network prediction model are as follows:
[0094] (1) Incorporating the sources of uncertainty and individual differences in the diesel engine assembly process into the neural network model, thereby obtaining confidence intervals for the prediction results and providing information on the accuracy of quality predictions;
[0095] (2) Set the dimensions of each layer of the Bayesian neural network: the input layer dimension is 39, the dimension of each of the three hidden layers is 40, and the output layer dimension is 1;
[0096] (3) Establish an uncertainty-aware Bayesian neural network prediction model. First, define a prior distribution p(ω i |β)~N(0,1), where ω i represents the weight of each process parameter, β represents the initial step set of process parameters, and then the posterior distribution of the model output is defined as Represents the parameter distribution of the network output, and the input assembly process parameters are represented as x i , the predicted final quality is
[0097] (4) Using the principle of variational inference, the variational posterior probability distribution p(ω i ) and the posterior distribution of the model output KL divergence of the difference value The final quality output of the Bayesian neural network is where f(x i ,ω i ) is the prediction function of the model;
[0098] (5) Develop uncertainty perception modules, such as Figure 2 As shown; the final quality output of the Bayesian neural network and the results of multiple forward propagations, the posterior distribution of the model output and the prior distribution p(ω i |β) between the deviation values;
[0099] (6) Calculate the confidence interval of the final quality of the prediction model using the expectation, variance and deviation values This statistical feature is further used to quantify the three uncertainties in the model architecture; adding residual perturbations ε~N(0,δ 2 I) quantify uncertainty in model architecture and data;
[0100] (7) Establish quantitative methods corresponding to the uncertainty of the excess distribution;
[0101] Defined as Where c is the total number of assembly features, m is the number of test samples, and y is the final quality value of the actual test;
[0102] (8) Establish a quantitative method for epistemic uncertainty;
[0103] Defined as The posterior discreteness of the parameters reflects the uncertainty of the quantitative model about the true value of the process parameters;
[0104] (9) Establish a method to quantify arbitrary uncertainty;
[0105] Defined as Where ε represents the residual variable, Represents the model output value of the residual perturbation; the corresponding quantitative distribution result is as follows Figure 4 shown.
[0106] Use SHAP interpretability analysis to associate uncertainty bounds with feature contributions, addressing the trade-off between model transparency and performance by providing confidence intervals for predictions and explicit feature contribution rankings.
[0107] Furthermore, we design a SHAP interpretability analysis module to associate uncertainty bounds with feature contributions. The specific steps are as follows:
[0108] (1) Define the numerical tensor of the forward propagation of the uncertainty-aware Bayesian neural network, and the prediction results of each diesel engine assembly data sample are stored in the form of a tensor;
[0109] (2) Use the SHAP interpretability analysis method to convert these tensors into parseable numerical values to ensure that the model output can be processed one by one;
[0110] (3) Quantifying the interaction between features, in the numerical tensors forward propagated by the Bayesian neural network with uncertainty-aware, the interaction effect is realized by slicing, combining and calculating the effect of feature combinations on the multi-dimensional tensor, specifically, τ is conditioned on the F dimension, where τ is a multi-dimensional tensor;
[0111] (4) Using SHAP explainability to analyze these tensors on a single subset of features, identifying the relevance of each feature's contribution to the output in different samples; (5) Defining the contribution of each assembly quality parameter Where, is defined as the global interaction score between assembly quality parameters and process features, f(R∪{x i}) as the contribution of the assembly quality feature to the global prediction, N represents the set of assembly quality parameters, R is a subset of the feature set,! represents factorial, and the contribution value of the assembly feature is as shown in Figure 4
[0112] (6) Defining the feature contribution interaction index, quantifying the contribution of feature interaction to the model output This index illustrates the way various features work together to affect the prediction result;
[0113] (7) Defining the local feature contribution correlation as quantifies the contribution of local feature interaction to the prediction uncertainty, and establishes an optimal relationship between the feature importance and the uncertainty boundary;
[0114] As shown in Figure 5 compared with other methods, the embodiment shows its performance advantage in quality prediction and uncertainty quantification.
[0115] The technical principles of the present application are described above in conjunction with specific embodiments, and these descriptions are only to explain the principles of the present application and cannot be interpreted in any way as a limitation on the scope of protection of the present application. Based on the explanations here, those skilled in the art can think of other specific embodiments of the present application without creative labor, and these ways will fall within the scope of protection of the present application.
Claims
1. A method for interpretable prediction of complex product assembly quality considering model uncertainty, characterized by: The steps include: S1. Collect historical data of complex product assembly process; S2, establishes an uncertainty-aware Bayesian neural network prediction module based on historical data; S3. Use a prediction model to characterize the mapping relationship between process parameters in the assembly process and the final quality indicators, and establish a Bayesian neural network uncertainty quantification module; S4, using SHAP interpretability to relate uncertainty bounds to feature contributions; S5. Based on the uncertainty-aware Bayesian neural network prediction module and the Bayesian neural network uncertainty quantification module, a global sensitivity analysis and a local interpretability module are constructed. The sensitivity analysis outputs the confidence of the predicted value, and the local interpretability outputs the contribution mapping relationship.
2. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 1, wherein S1 further comprises: Collect historical data of the diesel engine assembly process, perform data preprocessing on the historical data, and perform statistical processing on typical manufacturing process parameters; Among them, statistical processing includes calculating the mean, standard deviation and range of data characteristics.
3. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 1, The specific steps of establishing the uncertainty-aware Bayesian neural network prediction model in S2 include: S2.
1. Establish an uncertainty-aware Bayesian neural network prediction model. First, define a prior distribution p(ω i |β)~N(0,1), where ω i represents the weight of each process parameter, β represents the initial step set of process parameters, and then the posterior distribution of the model output is defined as Represents the parameter distribution of the network output, and the input assembly process parameters are represented as x i , the predicted final quality is S2.2, based on the prior distribution defined in S2.1 and the posterior distribution of the model output, using the principle of variational inference, the variational posterior probability distribution p(ω i ) and the posterior distribution of the model output KL divergence of the difference value The final quality output of the Bayesian neural network is where f(x i ,ω i ) is the prediction function of the model.
4. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 3, wherein S2 further comprises: S2.3, based on the final quality output of the Bayesian neural network in S2.2 and the results of multiple forward propagations, the posterior distribution of the model output and the prior distribution p(ω i The final quality and confidence interval of the prediction model are calculated using the expected, variance, and deviation values.
5. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 4, wherein S2 further comprises: S2.
4. Based on the predicted value and confidence interval of the final quality in S2.3, the final quality and assembly process parameters are correlated to achieve accurate prediction of the final assembly quality.
6. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 5, wherein S3 further comprises: S3.
1. Build a deep probabilistic architecture of an uncertainty-aware Bayesian neural network to output the final quality prediction value and the uncertainty ratio in the architecture. S3.
2. Establish a Bayesian neural network uncertainty quantification module and add residual perturbation ε~N(0,δ 2 I), quantifying uncertainty in model architecture and data.
7. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 6, wherein S3 further comprises: S3.2.
1. Develop quantitative methods corresponding to the uncertainty of the excess distribution; Defined as Where c is the total number of assembly features, m is the number of test samples, and y is the final quality value of the actual test; S3.2.
2. Establish a quantitative method for epistemic uncertainty; Defined as The posterior discreteness of the parameters reflects the uncertainty of the quantitative model about the true value of the process parameters; S3.2.
3. Establish a method to quantify arbitrary uncertainty; Defined as Where ε represents the residual variable, Represents the model output value of the residual perturbation.
8. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 1, wherein S4 further comprises: S4.
1. Define the contribution of each assembly quality parameter in, Defined as the global interaction score between assembly quality parameters and process characteristics, f(R∪{x i }) As the contribution of assembly quality features to global prediction, N represents the set of assembly quality parameters, R is the subset of the feature set, and ! represents the factorial; S4.
2. Define the local feature contribution correlation as Quantify the contribution of local feature interactions to prediction uncertainty and establish an optimal relationship between feature importance and uncertainty boundary.
9. The method for interpretable prediction of complex product assembly quality considering model uncertainty according to claim 1, further comprising: Establish an uncertainty quantification feedback mechanism, classify different uncertainties according to their sources, and dynamically adjust parameter weights based on the uncertainty distribution of assembly quality traceability; Apply the explainable prediction method of assembly quality to complex manufacturing systems, establish quantifiable process-quality relationships, and achieve accurate prediction of final quality indicators.