Product quality risk assessment method and system based on AI

Abstract

The invention discloses an AI-based product quality risk assessment method and system, and the method comprises the steps: obtaining the multi-source data of an automobile bearing production chain, carrying out the deep extraction, obtaining the multi-source features of the production chain, constructing a physical consistency penalty term, constructing a physically guided cross-modal attention matrix, carrying out the modal fusion, obtaining the state features of the production chain, and carrying out the prediction of the product quality risk. Constructing a manufacturing process dynamic directed graph and a risk propagation adjacency matrix, performing risk propagation to determine a risk value of each manufacturing process, and performing Bayesian optimization on a risk threshold value of each manufacturing process according to a bearing structure parameter and a bearing type to obtain a risk dynamic threshold value; and carrying out quality risk assessment on the automobile bearing according to the risk value and the risk dynamic threshold value of each manufacturing process. The method not only can improve the efficiency and accuracy of automobile bearing quality risk assessment, but also has good interpretability, and can be directly applied to an automobile bearing quality risk assessment system.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial product quality control technology, and in particular to an AI-based product quality risk assessment method and system. Background Technology

[0002] As a core component of automotive transmission systems, the quality and reliability of automotive bearings directly affect the safety and service life of the entire vehicle. With the automotive industry moving towards intelligent and lightweight development, the operating conditions of bearings are becoming increasingly stringent. The complex manufacturing process of automotive bearings involves multiple steps, including raw material procurement, turning, heat treatment, grinding, assembly, and testing. There are complex quality transfer and risk coupling relationships between these steps. How to accurately identify risk factors and assess product quality risks in real time from massive amounts of multi-source data has become a key technical challenge that restricts bearing manufacturers from improving their competitiveness.

[0003] However, existing technologies still have significant limitations: traditional machine learning methods are mostly based on a single data source and lack systematic modeling of physical factors such as raw material batch fluctuations and structural parameter tolerance coupling, resulting in a physical disconnect between feature extraction and risk assessment; at the same time, existing solutions mostly use static thresholds for single-station assessment, failing to establish a risk transmission model for the entire manufacturing process and unable to achieve cross-process potential defect early warning; furthermore, the determination of risk thresholds for different bearing models (such as wheel hub bearings and gearbox bearings) and different structural parameters relies on manual experience setting, lacking an adaptive optimization mechanism, which easily leads to missed inspection risks or excessive inspection costs. Therefore, this invention proposes an AI-based product quality risk assessment method and system. By constructing a physically guided cross-modal attention mechanism, building a manufacturing process risk propagation graph network, and Bayesian optimization of adaptive thresholds, it solves the problem of the lack of physical interpretability in traditional data-driven models, realizing a leap from single-node alarms to risk chain tracing, improving the accuracy and adaptability of risk assessment, and providing automotive bearing manufacturing enterprises with a comprehensive, accurate, and interpretable intelligent quality risk assessment solution. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based product quality risk assessment method and system.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Multi-source data of the automotive bearing production chain is acquired and deep extraction is performed to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; A physical consistency penalty term is constructed based on the multi-source features of the production chain. A physical-guided cross-modal attention matrix is ​​constructed from the multi-source features of the production chain and the physical consistency penalty term. Modal fusion is then performed to obtain the production chain state features. Data from the automotive bearing manufacturing process is extracted, and a dynamic directed graph and risk propagation adjacency matrix are constructed based on the production chain state characteristics. Risk propagation is then performed to determine the risk value of each manufacturing process. The manufacturing processes include turning, heat treatment, grinding, assembly, and inspection. The manufacturing process data includes process parameters and quality inspection results. Based on the bearing structural parameters and bearing type, Bayesian optimization is performed on the risk thresholds of each manufacturing process to obtain dynamic risk thresholds. Based on the risk values ​​of each manufacturing process and the dynamic risk thresholds, a quality risk assessment of the automotive bearing is conducted. The quality risk assessment includes an overall risk assessment and a manufacturing process risk assessment.

[0006] Furthermore, the method for performing deep extraction to obtain multi-source features of the production chain includes: Acquire multi-source data from the automotive bearing production chain; the multi-source data from the production chain includes raw material data, production line vibration signals, bearing structural parameters, and visual images; The measured quality parameters of each batch of raw materials are compared with the corresponding standards. Samples that do not meet the standards are selected to construct the batch quality residual tensor. The residual rate, quality deviation, and batch risk entropy of the corresponding batches are then calculated to form the raw material characteristics. The expression is as follows: ; ; ; in For batch quality residual tensor, for batch The index set of samples that did not meet the standards for batch The measured mass parameter vector of the raw material sample. This is the standard threshold vector for raw material quality parameters. For the bias mode mask, This refers to the batch raw material quality deviation. The number of dimensions for raw material quality parameters. For the first Risk sensitivity of each raw material quality parameter For the set of failure modes, For not meeting the standard The average values ​​of raw material quality parameters, , for The standard values ​​and allowable fluctuations of raw material quality parameters are specified in the standards. For batch risk entropy, for The probability of occurrence of a defect-like pattern. for Number of samples of class defect patterns This represents the total number of samples that did not meet the standards in the batch. Vibration signals from the production line at various time periods are collected and subjected to variational mode decomposition. Based on the bearing fault characteristic frequency theory, hard constraints are set to construct the variational mode decomposition optimization objective, expressed as: ; ; in The objective is to optimize the standard variational mode decomposition. For the first One eigenmode function The number of intrinsic mode functions. For the first One modal center frequency, for, For physical consistency loss, For physical consistency weights, For sparsity weights, For total variation weights, Eigenmode functions Total variation, For the set of fault characteristic frequencies, For fault characteristic frequency index, For the characteristic function, when the first... Each mode at frequency Power spectral density at Greater than the energy detection threshold Take 1 if the condition is met, otherwise take 0; The vibration characteristics of the production line are calculated based on the Hilbert spectral entropy and the fault energy ratio, center frequency, bandwidth, and spectral kurtosis of each mode of the vibration signal after variational mode decomposition optimization. The bearing structural parameters on the production line at different time periods are collected, and the sensitivity of each bearing structural parameter to contact stress and the probability of all parameters exceeding the functional limit are calculated to form the bearing structural parameter characteristics; the bearing structural parameters include elastic modulus, Poisson's ratio, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; Visual images of bearings on the production line at different times are collected. A ResNet-50 pre-trained neural network is used to perform defect detection, surface texture and color difference analysis on the bearing visual images. The bearing visual features are composed of defect ratio, defect distribution uniformity, texture features and metal color difference. The production chain multi-source features are composed of raw material features, production line vibration features, bearing structural parameter features and bearing visual features.

[0007] Furthermore, the method for obtaining the production chain state characteristics includes: A physical consistency penalty term is constructed based on the visual characteristics of the bearings and the vibration characteristics of the production line. A physical-guided cross-modal attention matrix is ​​then constructed from the multi-source characteristics of the production chain and the physical consistency penalty term, with the following expression: ; ; in For cross-modal attention weights, represent Bearing visual features Attention allocation coefficient for vibration characteristics of the production line. For visual feature vectors, For vibration characteristic vectors, To query the projection matrix, The key projection matrix, For physical constraint weights, As the key dimension, This is a penalty for physical consistency. for The center frequency of the frequency band The defect order is... The rotational frequency of the axis The diameter of the rolling element, The diameter of the pitch circle. Contact angle, For the outer ring characteristic function, for Defect location markings corresponding to visual characteristics of bearings For the outer raceway; The first fused state feature is obtained by modal fusion of bearing visual features, production line vibration features, and bearing structural parameter features. The expression is: ; in This is a characteristic of the first fusion state. The projection matrix is ​​the value. For layer normalization operation, Features of bearing structural parameters; A second fused state feature is obtained by fusing raw material features and bearing structural parameter features using a self-attention mechanism. The first fused state feature and the second fused state feature are then spliced ​​together to obtain the production chain state feature.

[0008] 5. Furthermore, the method for determining the risk value of each manufacturing process through risk propagation includes: Determine drawing nodes according to the automotive bearing manufacturing process. And determine the timing working status of each graph node. Determine the material flow relationships of each node in the diagram according to the automotive bearing manufacturing process. Extract the time-series process parameters and quality inspection results corresponding to each graph node, and combine them with the production chain status features to form graph node features. The dynamic directed graph of the manufacturing process is constructed. ,in, For manufacturing process quantity, The manufacturing process of the automotive bearings includes turning, heat treatment, grinding, assembly, and inspection; the process parameters include temperature, pressure, and rotational speed. The risk propagation strength between nodes in the dynamic directed graph of the manufacturing process is calculated as the edge weight between nodes. A risk propagation adjacency matrix is ​​constructed based on these edge weights, expressed as: ; in for Time map nodes For graph nodes The intensity of risk transmission, Due to process route constraints, For neighbor graph nodes Graph node features, Let be a joint probability distribution, representing the probability that all three states will occur simultaneously. Let be the conditional probability, representing the given graph nodes. Graph Nodes Current graph node features, future graph nodes Features are The probability, Let conditional probability represent the nodes of a known graph. Current graph node features, future graph nodes Features are The probability of; The risk propagation calculation is performed to determine the risk value of each node in the graph corresponding to the manufacturing process. The expression is as follows: ; in For graph nodes The risk value at the next moment. For graph nodes The upstream manufacturing process corresponds to the node risks in the graph. This represents the set of graph nodes corresponding to the upstream manufacturing process. This is the risk memory decay coefficient. For attention weights, The immune threshold for the manufacturing process.

[0009] Furthermore, the method for obtaining the dynamic risk threshold through Bayesian optimization includes: The prior distribution for Bayesian optimization is set based on the bearing structural parameters and bearing type. ,in, For the risk threshold vector, Represents the prior mean vector and prior covariance matrix It follows a Gaussian distribution; the prior mean vector includes the prior of the hub bearing and the prior of the gearbox bearing; the prior covariance matrix reflects the correlation of risk transmission between processes and is derived from the risk propagation adjacency matrix; The risk threshold adjustment strategy and data acquisition function are determined, and the dynamic risk threshold is obtained by performing Bayesian optimization to dynamically update the posterior distribution. The expression is as follows: ; ; ; in For acquisition functions, For historical risk threshold datasets, Candidate risk threshold vector The corresponding comprehensive risk cost function, The current optimal risk threshold The corresponding comprehensive risk cost function, To improve the quantity, Candidate risk threshold vector The corresponding false negative rate is less than the maximum acceptable false negative rate. The probability, , , As a risk cost weight, The false positive rate, To reduce the cost of early warning, for The dynamic risk threshold vector at any given time. For learning rate, For gradient operators, To explore noise.

[0010] Furthermore, the method for conducting quality risk assessment of automotive bearings includes: The risk value of each manufacturing process is compared with the corresponding dynamic risk threshold. When the risk value is greater than the corresponding dynamic risk threshold, the corresponding manufacturing process is determined to have manufacturing process risk. Calculate the average risk and the average dynamic risk threshold for all manufacturing processes, compare the data, and determine that there is an overall risk in the entire manufacturing process when the average risk is greater than the average dynamic risk threshold.

[0011] Secondly, an AI-based product quality risk assessment system includes: Feature extraction module: used to acquire multi-source data of the automotive bearing production chain and perform deep extraction to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; Feature fusion module: used to construct a physical consistency penalty term based on multi-source features of the production chain, construct a physically guided cross-modal attention matrix from the multi-source features of the production chain and the physical consistency penalty term, and perform modal fusion to obtain the production chain state features; Dynamic Directed Graph Module: Used to extract automotive bearing manufacturing process data, construct a dynamic directed graph of the manufacturing process and a risk propagation adjacency matrix by combining production chain state characteristics, and perform risk propagation to determine the risk value of each manufacturing process; the manufacturing process includes turning, heat treatment, grinding, assembly and inspection; the manufacturing process data includes process parameters and quality inspection results; Quality risk assessment module: used to perform Bayesian optimization on the risk thresholds of each manufacturing process based on bearing structural parameters and bearing type to obtain dynamic risk thresholds, and to conduct quality risk assessment of automotive bearings based on the risk values ​​and dynamic risk thresholds of each manufacturing process; the quality risk assessment includes overall risk assessment and manufacturing process risk assessment.

[0012] The beneficial effects of this invention are: This invention is an AI-based product quality risk assessment method and system. Compared with existing technologies, this invention has the following technical advantages: This invention overcomes the semantic gap problem caused by the simple splicing of multi-source data in traditional methods by deeply extracting and fusing the quality residuals of raw material batches, vibration signal modal features, structural parameter tolerance coupling information and visual defect features. This makes the feature associations learned by the model conform to the physical laws of bearing failure, and significantly improves the accuracy of defect identification and the generalization ability across working conditions. This invention constructs a dynamic directed graph and a risk propagation adjacency matrix with turning, heat treatment, grinding, assembly, and inspection as nodes, introduces a risk attenuation mechanism based on the physical characteristics of the process, establishes a quantitative model for the transmission of upstream process defects to downstream processes, realizes the transformation of quality control mode from post-screening to pre-prevention, and effectively reduces scrap rate and rework costs. This invention constructs a multi-objective optimization function that integrates false negative rate, false positive rate, and early warning cost for different bearing types and structural parameters. Through Bayesian optimization guided by the acquisition function and online gradient update mechanism, it automatically determines the optimal risk threshold for each process, significantly reducing the backlog cost of work-in-process caused by excessive inspection, and achieving precise quality risk control and maximizing economic benefits. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of an AI-based product quality risk assessment method according to the present invention. Detailed Implementation

[0014] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0015] The present invention provides an AI-based product quality risk assessment method and system, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Multi-source data of the automotive bearing production chain is acquired and deep extraction is performed to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; A physical consistency penalty term is constructed based on the multi-source features of the production chain. A physical-guided cross-modal attention matrix is ​​constructed from the multi-source features of the production chain and the physical consistency penalty term. Modal fusion is then performed to obtain the production chain state features. Data from the automotive bearing manufacturing process is extracted, and a dynamic directed graph and risk propagation adjacency matrix are constructed based on the production chain state characteristics. Risk propagation is then performed to determine the risk value of each manufacturing process. The manufacturing processes include turning, heat treatment, grinding, assembly, and inspection. The manufacturing process data includes process parameters and quality inspection results. Based on the bearing structural parameters and bearing type, Bayesian optimization is performed on the risk thresholds of each manufacturing process to obtain dynamic risk thresholds. Based on the risk values ​​of each manufacturing process and the dynamic risk thresholds, a quality risk assessment of the automotive bearing is conducted. The quality risk assessment includes an overall risk assessment and a manufacturing process risk assessment.

[0016] In this embodiment, the method for performing deep extraction to obtain multi-source features of the production chain includes: Acquire multi-source data from the automotive bearing production chain; the multi-source data from the production chain includes raw material data, production line vibration signals, bearing structural parameters, and visual images; The measured quality parameters of each batch of raw materials are compared with the corresponding standards. Samples that do not meet the standards are selected to construct the batch quality residual tensor. The residual rate, quality deviation, and batch risk entropy of the corresponding batches are then calculated to form the raw material characteristics. The expression is as follows: ; ; ; in For batch quality residual tensor, for batch The index set of samples that did not meet the standards for batch The measured mass parameter vector of the raw material sample. This is the standard threshold vector for raw material quality parameters. For the bias mode mask, This refers to the batch raw material quality deviation. The number of dimensions for raw material quality parameters. For the first Risk sensitivity of each raw material quality parameter For the set of failure modes, For not meeting the standard The average values ​​of raw material quality parameters, , for The standard values ​​and allowable fluctuations of raw material quality parameters are specified in the standards. For batch risk entropy, for The probability of occurrence of a defect-like pattern. for Number of samples of class defect patterns This represents the total number of samples that did not meet the standards in the batch. Vibration signals from the production line at various time periods are collected and subjected to variational mode decomposition. Based on the bearing fault characteristic frequency theory, hard constraints are set to construct the variational mode decomposition optimization objective, expressed as: ; ; in The objective of the standard variational mode decomposition optimization is... For the first One eigenmode function The number of intrinsic mode functions. For the first One modal center frequency, for, For physical consistency loss, For physical consistency weights, For sparsity weights, For total variation weights, Eigenmode functions Total variation, For the set of fault characteristic frequencies, For fault characteristic frequency index, For the characteristic function, when the first... Each mode at frequency Power spectral density at Greater than the energy detection threshold Take 1 if the condition is met, otherwise take 0; The vibration characteristics of the production line are calculated based on the Hilbert spectral entropy and the fault energy ratio, center frequency, bandwidth, and spectral kurtosis of each mode of the vibration signal after variational mode decomposition optimization. The bearing structural parameters on the production line at different time periods are collected, and the sensitivity of each bearing structural parameter to contact stress and the probability of all parameters exceeding the functional limit are calculated to form the bearing structural parameter characteristics; the bearing structural parameters include elastic modulus, Poisson's ratio, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; Visual images of bearings on the production line were collected at different times. ResNet-50 pre-trained neural network was used to perform defect detection, surface texture and color difference analysis on the bearing visual images. The bearing visual features were obtained by the defect ratio, defect distribution uniformity, texture features and metal color difference. The production chain multi-source features were composed of raw material features, production line vibration features, bearing structural parameter features and bearing visual features. In practical assessments, the risk sensitivity of raw material quality parameters is considered when calculating the degree of quality deviation. ,in The severity rating is the score indicating how severely the failure of this parameter affects the bearing's function; 10 indicates no warning or safety hazard, and 1 indicates no impact. Occurrence rate (this parameter is a score for the frequency of occurrence of the cause of failure, with 10 being very high (≥100 times / thousand pieces) and 1 being very low (≤0.01 times / thousand pieces). Detectability (a score of the current control measures’ ability to detect the failure of this parameter, where 10 is almost impossible to detect and 1 is almost certain to detect). In variational mode decomposition optimization objectives, the standard variational mode decomposition optimization objective expression is: ; in The original vibration signal, For time The partial derivatives, For Dirac delta function, For Hilbert transform kernel, The imaginary unit, For convolution operations, For complex exponential frequency shift operators, For fidelity weighting; This is a regularization term used to minimize the spectral bandwidth of each mode and promote spectral compactness of single modes. To generate an analytic signal, perform a Hilbert transform and construct the analytic signal. For analytic signals (complex signals, with a single-sided spectrum containing only positive frequency components), The demodulated signal gradient is used to convert the center frequency to a specific value. The mode shifts to the baseband, and its L2 norm measures the spectral bandwidth; To ensure fidelity, the sum of each mode approximates the original signal and controls reconstruction error; Bearing failure characteristic frequencies mainly include outer ring failure frequencies. Inner ring failure frequency Rolling element failure frequency Cage failure frequency The expression is: ; ; ; ; The outer ring failure frequency represents the impact frequency generated each time the rolling element passes a defect on the outer ring (applicable when the outer ring is stationary). The inner ring failure frequency represents the impact frequency generated each time the rolling element passes a defect on the inner ring (the frequency is slightly higher when the inner ring is rotating). The rolling element failure frequency represents the frequency at which the defect contacts both the inner and outer rings once per revolution of the rolling element (usually 2-4 times the shaft rotation frequency). The cage failure frequency is usually 0.4-0.5 times the shaft rotation frequency. The rotational frequency of the axis For the number of rolling elements, The diameter of the rolling element, The diameter of the pitch circle. Contact angle; The expressions for the sensitivity of each bearing structural parameter to contact stress and the probability of all parameters jointly exceeding the functional limit are as follows: ; ; in Bearing structural parameters Contact stress Sensitivity , , For the Hertzian contact ellipse semi-axis, For rolling element loads, Given the probability that all parameters exceed the functional limit, all bearing structural parameters follow a truncated normal distribution. For the failure domain, To meet the lifespan requirement, Bearing structural parameters Corresponding actual lifespan, Pick , For the basic rated life, For equivalent dynamic load, Life expectancy index For the basic rated dynamic load, The material coefficient for bearing steel is 1.3-1.8. This is the process coefficient. The number of columns for the rolling element; the multi-source characteristics of the production chain are generated according to each time period and production line.

[0017] In this embodiment, the method for obtaining production chain state characteristics includes: A physical consistency penalty term is constructed based on the visual characteristics of the bearings and the vibration characteristics of the production line. A physical-guided cross-modal attention matrix is ​​then constructed from the multi-source characteristics of the production chain and the physical consistency penalty term, with the following expression: ; ; in For cross-modal attention weights, represent Bearing visual features Attention allocation coefficient for vibration characteristics of the production line. For visual feature vectors, For vibration characteristic vectors, To query the projection matrix, The key projection matrix, For physical constraint weights, As the key dimension, This is a penalty for physical consistency. for The center frequency of the frequency band The defect order is... The rotational frequency of the axis The diameter of the rolling element, The diameter of the pitch circle. Contact angle, For the outer ring characteristic function, for Defect location markings corresponding to visual characteristics of bearings For the outer raceway; The first fused state feature is obtained by modal fusion of bearing visual features, production line vibration features, and bearing structural parameter features. The expression is: ; in This is a characteristic of the first fusion state. The projection matrix is ​​the value. For layer normalization operation, Features of bearing structural parameters; A second fused state feature is obtained by fusing raw material features and bearing structural parameter features using a self-attention mechanism. The first fused state feature and the second fused state feature are then spliced ​​together to obtain the production chain state feature. In practical evaluation, in the physical consistency penalty term, the outer ring indicator function is used when the defect is located on the outer ring. At this time, the vibration frequency band is required to be... It must be close to the nth harmonic of the outer ring fault characteristic frequency; otherwise, a large penalty is imposed. When the defect is not located on the outer ring, the outer ring characteristic function... At this point, the penalty term degenerates into a regularization of the frequency itself; Physical constraint weights Take 0.5, defect order Set the value to 1, and the attention dimension to 128.

[0018] In this embodiment, the method for determining the risk value of each manufacturing process through risk propagation includes: Determine drawing nodes according to the automotive bearing manufacturing process. And determine the timing working status of each graph node. Determine the material flow relationships of each node in the diagram according to the automotive bearing manufacturing process. Extract the time-series process parameters and quality inspection results corresponding to each graph node, and combine them with the production chain status features to form graph node features. The dynamic directed graph of the manufacturing process is constructed. ,in, For manufacturing process quantity, The manufacturing process of the automotive bearings includes turning, heat treatment, grinding, assembly, and inspection; the process parameters include temperature, pressure, and rotational speed. The risk propagation strength between nodes in the dynamic directed graph of the manufacturing process is calculated as the edge weight between nodes. A risk propagation adjacency matrix is ​​constructed based on these edge weights, expressed as: ; in for Time map nodes For graph nodes The intensity of risk transmission, Due to process route constraints, For neighbor graph nodes Graph node features, Let be a joint probability distribution, representing the probability that all three states will occur simultaneously. Let be the conditional probability, representing the given graph nodes. Graph Nodes Current graph node features, future graph nodes Features are The probability, Let conditional probability represent the nodes of a known graph. Current graph node features, future graph nodes Features are The probability of; The risk propagation calculation is performed to determine the risk value of each node in the graph corresponding to the manufacturing process. The expression is as follows: ; in For graph nodes The risk value at the next moment. For graph nodes The upstream manufacturing process corresponds to the node risks in the graph. This represents the set of graph nodes corresponding to the upstream manufacturing process. This is the risk memory decay coefficient. For attention weights, The immune threshold for the manufacturing process; In actual assessments, the automotive bearing manufacturing process specifically follows the sequence of "turning → heat treatment → grinding → assembly → inspection," constrained by the process route. This indicates that material transfer is permitted only between workstations where there is actual material flow (e.g., "turning" only has a material flow relationship with "heat treatment", and "heat treatment" has a material flow relationship with "turning" and "grinding"). Process immunity threshold of heat treatment process The risk value of the "turning" process is relatively high and can repair previous defects. The risk value of the "turning" process is determined based on the quality inspection results of the production chain in the previous period (the product of the defect category score, the reciprocal of the quality grade, and the confidence level of the inspection results). The risk values ​​of the remaining manufacturing processes are calculated based on the risk value of the "turning" process through the above risk propagation. The quality grade is divided by historical bearing defect samples, mainly including dimensional defects (inner diameter, outer diameter, width), shape defects (roundness, cylindricity, perpendicularity), surface defects (surface roughness, vibration marks, scratches), and material defects (cracks, folds, inclusions).

[0019] In this embodiment, the method for obtaining a dynamic risk threshold through Bayesian optimization includes: The prior distribution for Bayesian optimization is set based on the bearing structural parameters and bearing type. ,in, For the risk threshold vector, Represents the prior mean vector and prior covariance matrix It follows a Gaussian distribution; the prior mean vector includes the prior of the hub bearing and the prior of the gearbox bearing; the prior covariance matrix reflects the correlation of risk transmission between processes and is derived from the risk propagation adjacency matrix; The risk threshold adjustment strategy and data acquisition function are determined, and the dynamic risk threshold is obtained by performing Bayesian optimization to dynamically update the posterior distribution. The expression is as follows: ; ; ; in For acquisition functions, For historical risk threshold datasets, Candidate risk threshold vector The corresponding comprehensive risk cost function, The current optimal risk threshold The corresponding comprehensive risk cost function, To improve the quantity, Candidate risk threshold vector The corresponding false negative rate is less than the maximum acceptable false negative rate. The probability, , , As a risk cost weight, The false positive rate, To reduce the cost of early warning, for The dynamic risk threshold vector at any given time. For learning rate, For gradient operators, To explore noise; In actual evaluation, the prior mean vector includes the wheel hub bearing prior (high reliability requirements, overall low threshold) and the gearbox bearing prior (high-speed operating conditions, particularly sensitive to the grinding process threshold), and is taken according to the entire automotive bearing manufacturing process. , ; In the comprehensive risk cost function In the context of the false negative rate function The false positive rate function represents the probability that a quality risk actually exists but is not identified as a high risk by the system (false negative rate). The false alarm rate (false positive rate) represents the probability that a system mistakenly classifies a situation as high-risk when there is no actual quality risk. Its expression is: ; ; in This represents the number of failed samples. For the set of failed samples, For indicator functions, The time when the failed sample occurred. for time The sample corresponds to the predicted risk value of the manufacturing process. for time Risk threshold of the manufacturing process for Sample corresponds to manufacturing process index. This represents the normal sample size. for The sample corresponds to the predicted risk value of the manufacturing process. for Risk threshold of the manufacturing process for Sample corresponds to manufacturing process index. for The false alarm cost coefficient for the manufacturing process is set to 1 / 3 / 2 / 5 for turning / heat treatment / grinding / assembly respectively. The scrap cost of the bearing with current structural parameters. This is the average cost of scrapping; Early warning costs It consists of increased inspection costs (linearly correlated with advance warning time), work-in-process inventory costs, lifetime improvement benefits, and failure prevention benefits (take 1- ); Explore noise It follows a multivariate normal distribution (isotropic). To explore variance.

[0020] In this embodiment, the method for conducting a quality risk assessment of automotive bearings includes: The risk value of each manufacturing process is compared with the corresponding dynamic risk threshold. When the risk value is greater than the corresponding dynamic risk threshold, the corresponding manufacturing process is determined to have manufacturing process risk. Calculate the average risk and the average dynamic risk threshold for all manufacturing processes, compare the data, and determine that there is an overall risk in the entire manufacturing process when the average risk is greater than the average dynamic risk threshold. In actual assessment, it is also necessary to compare the characteristics of raw materials with the corresponding characteristic thresholds, and conduct risk assessment of raw material data based on the comparison results to facilitate the traceability of raw material types with potential risks.

[0021] Secondly, an AI-based product quality risk assessment system includes: Feature extraction module: used to acquire multi-source data of the automotive bearing production chain and perform deep extraction to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; Feature fusion module: used to construct a physical consistency penalty term based on multi-source features of the production chain, construct a physically guided cross-modal attention matrix from the multi-source features of the production chain and the physical consistency penalty term, and perform modal fusion to obtain the production chain state features; Dynamic Directed Graph Module: Used to extract automotive bearing manufacturing process data, construct a dynamic directed graph of the manufacturing process and a risk propagation adjacency matrix by combining production chain state characteristics, and perform risk propagation to determine the risk value of each manufacturing process; the manufacturing process includes turning, heat treatment, grinding, assembly and inspection; the manufacturing process data includes process parameters and quality inspection results; Quality risk assessment module: used to perform Bayesian optimization on the risk thresholds of each manufacturing process based on bearing structural parameters and bearing type to obtain dynamic risk thresholds, and to conduct quality risk assessment of automotive bearings based on the risk values ​​and dynamic risk thresholds of each manufacturing process; the quality risk assessment includes overall risk assessment and manufacturing process risk assessment.

[0022] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A product quality risk assessment method based on AI, characterized in that, Includes the following steps: S1. Obtain multi-source data of the automotive bearing production chain and perform deep extraction to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; S2. Construct a physical consistency penalty term based on the multi-source characteristics of the production chain. Construct a physical-guided cross-modal attention matrix based on the multi-source characteristics of the production chain and the physical consistency penalty term. Perform modal fusion to obtain the production chain state characteristics. S3. Extract automotive bearing manufacturing process data, combine production chain state characteristics to construct a dynamic directed graph of the manufacturing process and a risk propagation adjacency matrix, and perform risk propagation to determine the risk value of each manufacturing process. The manufacturing process includes turning, heat treatment, grinding, assembly, and inspection; The manufacturing process data includes process parameters and quality inspection results; S4. Based on the bearing structural parameters and bearing type, Bayesian optimization is performed on the risk thresholds of each manufacturing process to obtain dynamic risk thresholds. Based on the risk values ​​of each manufacturing process and the dynamic risk thresholds, a quality risk assessment of the automotive bearing is conducted. The quality risk assessment includes an overall risk assessment and a manufacturing process risk assessment.

2. The AI-based product quality risk assessment method according to claim 1, characterized in that, The method for obtaining multi-source features of the production chain through deep extraction includes: Acquire multi-source data from the automotive bearing production chain; the multi-source data from the production chain includes raw material data, production line vibration signals, bearing structural parameters, and visual images; The measured quality parameters of each batch of raw materials are compared with the corresponding standards. Samples that do not meet the standards are selected to construct the batch quality residual tensor. The residual rate, quality deviation, and batch risk entropy of the corresponding batches are then calculated to form the raw material characteristics. The expression is as follows: ； ； ； in For batch quality residual tensor, for batch The index set of samples that did not meet the standards for batch The measured mass parameter vector of the raw material sample. This is the standard threshold vector for raw material quality parameters. For the bias mode mask, This refers to the batch raw material quality deviation. The number of dimensions for raw material quality parameters. For the first Risk sensitivity of each raw material quality parameter For the set of failure modes, For not meeting the standard The average values ​​of raw material quality parameters, , for The standard values ​​and allowable fluctuations of raw material quality parameters are specified in the standards. For batch risk entropy, for The probability of occurrence of a defect-like pattern. for Number of samples of defect patterns This represents the total number of samples that did not meet the standards in the batch. Vibration signals from the production line at various time periods are collected and subjected to variational mode decomposition. Based on the bearing fault characteristic frequency theory, hard constraints are set to construct the variational mode decomposition optimization objective, expressed as: ； ； in The objective of the standard variational mode decomposition optimization is... For the first One eigenmode function The number of intrinsic mode functions. For the first One modal center frequency, for, For physical consistency loss, For physical consistency weights, For sparsity weights, For total variation weights, Eigenmode functions Total variation, For the set of fault characteristic frequencies, For fault characteristic frequency index, For the characteristic function, when the first... Each mode at frequency Power spectral density at Greater than the energy detection threshold Take 1 if the condition is met, otherwise take 0; The vibration characteristics of the production line are calculated based on the Hilbert spectral entropy and the fault energy ratio, center frequency, bandwidth, and spectral kurtosis of each mode of the vibration signal after variational mode decomposition optimization. The bearing structural parameters on the production line at different time periods are collected, and the sensitivity of each bearing structural parameter to contact stress and the probability of all parameters exceeding the functional limit are calculated to form the bearing structural parameter characteristics; the bearing structural parameters include elastic modulus, Poisson's ratio, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; Visual images of bearings on the production line at different times are collected. A ResNet-50 pre-trained neural network is used to perform defect detection, surface texture and color difference analysis on the bearing visual images. The bearing visual features are composed of defect ratio, defect distribution uniformity, texture features and metal color difference. The production chain multi-source features are composed of raw material features, production line vibration features, bearing structural parameter features and bearing visual features.

3. The AI-based product quality risk assessment method according to claim 1, characterized in that, The method for obtaining production chain state characteristics includes: A physical consistency penalty term is constructed based on the visual characteristics of the bearings and the vibration characteristics of the production line. A physical-guided cross-modal attention matrix is ​​then constructed from the multi-source characteristics of the production chain and the physical consistency penalty term, with the following expression: ； ； in For cross-modal attention weights, represent Bearing visual features Attention allocation coefficient for vibration characteristics of the production line. For visual feature vectors, For vibration characteristic vectors, To query the projection matrix, The key projection matrix, For physical constraint weights, As the key dimension, This is a penalty for physical consistency. for The center frequency of the frequency band The defect order is... The rotational frequency of the axis The diameter of the rolling element, The diameter of the pitch circle. Contact angle, For the outer ring characteristic function, for Defect location markings corresponding to visual characteristics of bearings For the outer raceway; Modal fusion of bearing visual features, production line vibration features, and bearing structural parameter features yields the first fused state feature, expressed as: ； in This is a characteristic of the first fusion state. The projection matrix is ​​the value. For layer normalization operation, Features of bearing structural parameters; A second fused state feature is obtained by fusing raw material features and bearing structural parameter features using a self-attention mechanism. The first fused state feature and the second fused state feature are then spliced ​​together to obtain the production chain state feature.

4. The AI-based product quality risk assessment method according to claim 1, characterized in that, The method for determining the risk value of each manufacturing process through risk propagation includes: Determine drawing nodes according to the automotive bearing manufacturing process. And determine the timing working status of each graph node. Determine the material flow relationships of each node in the diagram according to the automotive bearing manufacturing process. Extract the time-series process parameters and quality inspection results corresponding to each graph node, and combine them with the production chain status features to form graph node features. The dynamic directed graph of the manufacturing process is constructed. ,in, For manufacturing process quantity, The manufacturing process of the automotive bearings includes turning, heat treatment, grinding, assembly, and inspection; the process parameters include temperature, pressure, and rotational speed. The risk propagation strength between nodes in the dynamic directed graph of the manufacturing process is calculated as the edge weight between nodes. A risk propagation adjacency matrix is ​​constructed based on these edge weights, expressed as: ； in for Time map nodes For graph nodes The intensity of risk transmission, Due to process route constraints, For neighbor graph nodes Graph node features, Let be a joint probability distribution, representing the probability that all three states will occur simultaneously. Let be the conditional probability, representing the given graph nodes. Graph Nodes Current graph node features, future graph nodes Features are The probability, Let conditional probability represent the nodes of a known graph. Current graph node features, future graph nodes Features are The probability of; The risk propagation calculation is performed to determine the risk value of each node in the graph corresponding to the manufacturing process. The expression is as follows: ； in For graph nodes The risk value at the next moment. For graph nodes The upstream manufacturing process corresponds to the node risks in the graph. This represents the set of graph nodes corresponding to the upstream manufacturing process. This is the risk memory decay coefficient. For attention weights, The immune threshold for the manufacturing process.

5. The AI-based product quality risk assessment method according to claim 1, characterized in that, The method for obtaining the dynamic risk threshold through Bayesian optimization includes: The prior distribution for Bayesian optimization is set based on the bearing structural parameters and bearing type. ,in, For the risk threshold vector, Represents the prior mean vector and prior covariance matrix It follows a Gaussian distribution; the prior mean vector includes the prior of the hub bearing and the prior of the gearbox bearing; the prior covariance matrix reflects the correlation of risk transmission between processes and is derived from the risk propagation adjacency matrix; The risk threshold adjustment strategy and data acquisition function are determined, and the dynamic risk threshold is obtained by performing Bayesian optimization to dynamically update the posterior distribution. The expression is as follows: ； ； ； in For acquisition functions, For historical risk threshold datasets, Candidate risk threshold vector The corresponding comprehensive risk cost function, The current optimal risk threshold The corresponding comprehensive risk cost function, To improve the quantity, Candidate risk threshold vector The corresponding false negative rate is less than the maximum acceptable false negative rate. The probability, , , As a risk cost weight, The false positive rate, To reduce the cost of early warning, for The dynamic risk threshold vector at any given time. For learning rate, For gradient operators, To explore noise.

6. The AI-based product quality risk assessment method according to claim 1, characterized in that, The method for conducting quality risk assessment of automotive bearings includes: The risk value of each manufacturing process is compared with the corresponding dynamic risk threshold. When the risk value is greater than the corresponding dynamic risk threshold, the corresponding manufacturing process is determined to have manufacturing process risk. Calculate the average risk and the average dynamic risk threshold for all manufacturing processes, compare the data, and determine that there is an overall risk in the entire manufacturing process when the average risk is greater than the average dynamic risk threshold.

7. An AI-based product quality risk assessment system, used to perform the method described in any one of claims 1-6, characterized in that, include: Feature extraction module: used to acquire multi-source data of the automotive bearing production chain and perform deep extraction to obtain multi-source features of the production chain; the multi-source data of the production chain includes raw material data, production line vibration signals, bearing structural parameters and visual images; Feature fusion module: used to construct a physical consistency penalty term based on multi-source features of the production chain, construct a physically guided cross-modal attention matrix from the multi-source features of the production chain and the physical consistency penalty term, and perform modal fusion to obtain the production chain state features; Dynamic Directed Graph Module: Used to extract automotive bearing manufacturing process data, combine production chain state characteristics to construct a dynamic directed graph of the manufacturing process and a risk propagation adjacency matrix, and perform risk propagation to determine the risk value of each manufacturing process; The manufacturing process includes turning, heat treatment, grinding, assembly, and inspection; The manufacturing process data includes process parameters and quality inspection results; Quality risk assessment module: used to perform Bayesian optimization on the risk thresholds of each manufacturing process based on bearing structural parameters and bearing type to obtain dynamic risk thresholds, and to conduct quality risk assessment of automotive bearings based on the risk values ​​and dynamic risk thresholds of each manufacturing process; the quality risk assessment includes overall risk assessment and manufacturing process risk assessment.

Images