CN-122020485-A - Product quality risk assessment method and system based on AI

CN122020485ACN 122020485 ACN122020485 ACN 122020485ACN-122020485-A

Abstract

The invention discloses a product quality risk assessment method and system based on AI, comprising the steps of obtaining multi-source data of an automobile bearing production chain, carrying out deep extraction to obtain multi-source characteristics of the production chain, constructing a physical consistency penalty item, constructing a physical guidance cross-modal attention matrix, carrying out modal fusion to obtain state characteristics of the production chain, constructing a dynamic directed graph and a risk propagation adjacent matrix of a manufacturing process, carrying out risk propagation to determine risk values of all the manufacturing processes, carrying out Bayesian optimization on the risk thresholds of all the manufacturing processes according to bearing structure parameters and bearing types to obtain risk dynamic thresholds, and carrying out quality risk assessment of the automobile bearing according to the risk values and the risk dynamic thresholds of all the manufacturing processes. The method not only can improve the efficiency and accuracy of the quality risk assessment of the automobile bearing, but also has better interpretability, and can be directly applied to an automobile bearing quality risk assessment system.

Inventors

WEI FANGFANG
ZHAO HAIGANG
YANG QINGPING
ZHU ZHIHAO
ZHOU KEYU

Assignees

万向钱潮股份公司

Dates

Publication Date: 20260512
Application Date: 20260226

Claims (7)

1. An AI-based product quality risk assessment method, comprising the steps of: S1, acquiring multi-source data of an automobile bearing production chain, and performing deep extraction to obtain multi-source characteristics of the production chain, wherein the multi-source data of the production chain comprises raw material data, production line vibration signals, bearing structure parameters and visual images; S2, constructing a physical consistency penalty item according to the multi-source characteristics of the production chain, constructing a physically guided cross-modal attention matrix by the multi-source characteristics of the production chain and the physical consistency penalty item, and carrying out modal fusion to obtain the state characteristics of the production chain; S3, extracting automobile bearing manufacturing process data, constructing a dynamic directed graph of the manufacturing process and a risk propagation adjacent matrix by combining production chain state characteristics, and performing risk propagation to determine risk values of all manufacturing processes, wherein the manufacturing processes comprise turning, heat treatment, grinding, assembly and detection; S4, performing Bayesian optimization on the risk threshold value of each manufacturing process according to the bearing structure parameters and the bearing type to obtain a risk dynamic threshold value, and performing quality risk assessment of the automobile bearing according to the risk value and the risk dynamic threshold value of each manufacturing process, wherein the quality risk assessment comprises overall risk assessment and manufacturing process risk assessment.
2. The AI-based product quality risk assessment method of claim 1, wherein the method for deep extraction to obtain production chain multi-source features comprises: the method comprises the steps of obtaining multi-source data of an automobile bearing production chain, wherein the multi-source data of the production chain comprises raw material data, production line vibration signals, bearing structure parameters and visual images; Comparing measured quality parameters of raw materials of each batch with corresponding standards, screening unqualified samples to construct a batch quality residual tensor, and calculating the residual rate, quality deviation degree and batch risk entropy of the corresponding batch to form raw material characteristics, wherein the expression is as follows: ; ; ; Wherein the method comprises the steps of As a batch quality residual tensor, Is that Batch of The index set of samples not up to standard in the process, Is that Batch of The measured quality parameter vector of the raw material sample, Is a standard threshold vector of the quality parameters of the raw materials, For the deviation pattern mask to be used, Is the quality deviation degree of batch raw materials, Is the dimension number of the quality parameters of the raw materials, Is the first Risk sensitivity of individual raw material quality parameters, In order to be a set of failure modes, Is not up to standard The average value of the quality parameters of the raw materials, 、 Is that The standard specification value and standard allowable fluctuation of the quality parameters of the raw materials, For the entropy of the batch risk, Is that The probability of occurrence of a defect-like pattern, Is that The number of samples of the defect-like pattern, The total number of the samples which are not up to standard in the batch is calculated; vibration signals on the production line at each period are collected to carry out variation modal decomposition, a hard constraint construction variation modal decomposition optimization target is set according to a bearing fault characteristic frequency theory, and the expression is as follows: ; ; Wherein the method comprises the steps of The target is optimized for standard variant modal decomposition, Is the first The number of eigenmode functions, As the number of eigenmode functions, Is the first The center frequency of the individual modes, In order to achieve this, the first and second, In the event of a loss of physical consistency, For the weight of the physical consistency, For the sparsity weight to be weighted, For the total variation of the weight of the component, As an eigenmode function Is characterized by that the total variation of said (a) is, As a set of failure feature frequencies, For the failure feature frequency index, As a sexual function, when The frequency of each mode Power spectral density at Greater than the energy detection threshold Taking 1 when the time is, otherwise taking 0; Calculating Hilbert spectrum entropy and fault energy duty ratio, center frequency, bandwidth and spectrum kurtosis of each mode according to vibration signals after decomposition optimization of the variation modes to form a production line vibration characteristic; The method comprises the steps of collecting bearing structure parameters on a production line at each time period, and calculating the sensitivity of each bearing structure parameter to contact stress and the probability that all parameters are combined to exceed the functional limit to form bearing structure parameter characteristics, wherein the bearing structure parameters comprise elastic modulus, poisson ratio, pitch diameter, rolling body quantity and contact angle; And acquiring bearing visual images on a production line at each time interval, performing defect detection, surface texture and color difference analysis on the bearing visual images by adopting ResNet-50 pre-training neural networks to obtain a bearing visual characteristic composed of defect duty ratio, defect distribution uniformity, texture characteristics and metal color difference, and forming a production chain multisource characteristic by raw material characteristics, production line vibration characteristics, bearing structure parameter characteristics and bearing visual characteristics.
3. The AI-based product quality risk assessment method of claim 1, wherein the method of obtaining production chain status characteristics comprises: constructing a physical consistency penalty term according to the visual characteristics of the bearing and the vibration characteristics of the production line, and constructing a physically guided cross-modal attention matrix by the multisource characteristics of the production chain and the physical consistency penalty term, wherein the expression is as follows: ; ; Wherein the method comprises the steps of For cross-modal attention weights, represent Bearing visual characteristic pair Attention distribution coefficient of vibration characteristics of the production line, As a vector of the visual characteristics, As a feature vector of the vibration, In order to query the projection matrix, For the key projection matrix, For the weight of the physical constraint, In order to be a dimension of a key, For the physical consistency penalty term, Is that The center frequency of the frequency band, In order to be a defect order, For the frequency of rotation of the shaft, For the diameter of the rolling bodies, In order to achieve the pitch diameter, As a contact angle of the glass, As an outer circle of the indication function, Is that The defect position mark corresponding to the visual characteristic of the bearing, Is an outer ring raceway; carrying out modal fusion on the visual characteristics of the bearing, the vibration characteristics of the production line and the structural parameter characteristics of the bearing to obtain a first fusion state characteristic, wherein the expression is as follows: ; Wherein the method comprises the steps of As a feature of the first fusion state, For the projection matrix of the values, For the layer normalization operation, Is a bearing structure parameter characteristic; And fusing the raw material characteristics and the bearing structure parameter characteristics by adopting a self-attention mechanism to obtain second fused state characteristics, and splicing the first fused state characteristics and the second fused state characteristics to obtain production chain state characteristics.
4. The AI-based product quality risk assessment method of claim 1, wherein the method of risk propagation determining risk values for each manufacturing process comprises: determining a graph node according to a manufacturing process of a vehicle bearing And determining the time sequence working state of each graph node Determining the material circulation relation of each graph node according to the manufacturing flow of the automobile bearing Extracting time sequence technological parameters and quality detection results of the manufacturing process corresponding to each graph node and forming graph node characteristics with production chain state characteristics Dynamic directed graph by build manufacturing process , wherein, In order to make the number of processes to be performed, The manufacturing process of the automobile bearing comprises turning, heat treatment, grinding, assembly and detection, wherein the technological parameters comprise temperature, pressure and rotating speed; Calculating the risk conduction intensity among nodes of each graph of the dynamic directed graph in the manufacturing process as the edge weight among the nodes of each graph, and constructing a risk propagation adjacent matrix according to the edge weight among the nodes of each graph, wherein the expression is as follows: ; Wherein the method comprises the steps of Is that Time graph node Opposite graph node Is used for the risk of conducting the intensity of the risk, For the purposes of the process route constraints, Is a neighbor graph node Is characterized by the graph node characteristics of (c), For joint probability distribution, the probability of the simultaneous occurrence of three states is represented, For conditional probability, representing known graph nodes Sum graph node Current moment diagram node characteristics, future diagram node Is characterized in that Is a function of the probability of (1), Representing known graph nodes for conditional probability Current moment diagram node characteristics, future diagram node Is characterized in that Probability of (2); and carrying out risk propagation to calculate a risk value of each graph node corresponding to the manufacturing process, wherein the expression is as follows: ; Wherein the method comprises the steps of Is a graph node The risk value at the next moment in time, Is a graph node The upstream manufacturing process of (c) corresponds to the graph node risk, For an upstream manufacturing process map node set, As a factor of the attenuation of the risk memory, In order for the attention to be weighted, Immune thresholds are the manufacturing process.
5. The AI-based product quality risk assessment method of claim 1, wherein the bayesian optimization method for obtaining the risk dynamic threshold comprises: setting a priori distribution of Bayesian optimization according to bearing structure parameters and bearing types , wherein, As a risk threshold value vector, Representing a priori mean vectors A priori covariance matrix The prior mean value vector comprises a hub bearing prior and a gearbox bearing prior, wherein the prior covariance matrix reflects the correlation of risk transfer among working procedures and is derived from a risk propagation adjacent matrix; Determining a risk threshold adjustment strategy and an acquisition function, and performing Bayesian optimization dynamic update posterior distribution to obtain a risk dynamic threshold, wherein the expression is: ; ; ; Wherein the method comprises the steps of In order to collect the function of the object, As a historical risk threshold data set, For candidate risk threshold vectors A corresponding integrated risk cost function is provided, Is the current optimal risk threshold A corresponding integrated risk cost function is provided, In order to improve the quality of the product, For candidate risk threshold vectors The corresponding false negative rate is smaller than the maximum acceptable false negative rate Is a function of the probability of (1), 、、 As a weight of the risk cost, For a false positive rate, the positive rate, In order to early warn the cost of the vehicle, Is that The risk dynamic threshold vector for the moment in time, In order for the rate of learning to be high, For the purpose of the gradient operator, To explore the noise.
6. The AI-based product quality risk assessment method of claim 1, wherein the method for quality risk assessment of automotive bearings comprises: Comparing the risk value of each manufacturing process with a corresponding risk dynamic threshold value, and judging that the manufacturing process risk exists in the corresponding manufacturing process when the risk value is larger than the corresponding risk dynamic threshold value; And calculating the risk average value and the risk dynamic threshold value average value of all the manufacturing processes, comparing data, and judging that the whole risk exists in the whole manufacturing process when the risk average value is larger than the risk dynamic threshold value average value.
7. AI-based product quality risk assessment system for performing the method of any of claims 1-6, comprising: The device comprises a characteristic extraction module, a characteristic extraction module and a control module, wherein the characteristic extraction module is used for obtaining multi-source data of an automobile bearing production chain and carrying out deep extraction to obtain multi-source characteristics of the production chain, and the multi-source data of the production chain comprises raw material data, production line vibration signals, bearing structure parameters and visual images; The feature fusion module is used for constructing a physical consistency penalty item according to the multi-source features of the production chain, constructing a physically guided cross-modal attention matrix by the multi-source features of the production chain and the physical consistency penalty item, and carrying out modal fusion to obtain the state features of the production chain; the dynamic directed graph module is used for extracting the data of the manufacturing process of the automobile bearing, constructing a dynamic directed graph and a risk transmission adjacent matrix of the manufacturing process by combining the state characteristics of a production chain, and carrying out risk transmission to determine the risk value of each manufacturing process, wherein the manufacturing process comprises turning, heat treatment, grinding, assembly and detection; The quality risk assessment module is used for carrying out Bayesian optimization on the risk threshold value of each manufacturing process according to the bearing structure parameters and the bearing type to obtain a risk dynamic threshold value, and carrying out quality risk assessment of the automobile bearing according to the risk value and the risk dynamic threshold value of each manufacturing process, wherein the quality risk assessment comprises overall risk assessment and manufacturing process risk assessment.

Description

Product quality risk assessment method and system based on AI Technical Field The invention relates to the technical field of intelligent manufacturing and industrial product quality control, in particular to a product quality risk assessment method and system based on AI. Background The automobile bearing is used as a core basic part of an automobile transmission system, the quality reliability of the automobile bearing is directly related to the safety and the service life of the whole automobile, along with the development of the automobile industry to the intelligent and light weight directions, the service working condition of the bearing is increasingly harsh, the automobile bearing is subjected to complex manufacturing processes such as raw material purchase, turning, heat treatment, grinding, assembly and detection, complex quality transmission and risk coupling relations exist among the processes, and how to accurately identify risk factors from massive multi-source data and evaluate the quality risk of products in real time becomes a key technical problem for restricting bearing manufacturing enterprises to promote competitiveness. However, the prior art still has significant limitations that the traditional machine learning method is mostly based on a single data source, lacks systematic modeling of physical factors such as raw material batch fluctuation, structural parameter tolerance coupling and the like, and leads to physical disconnection of feature extraction and risk judgment; in addition, the determination of the risk threshold values for different types of bearings (such as a hub bearing and a gearbox bearing) and different structural parameters depends on manual experience setting, and the lack of an adaptive optimization mechanism is easy to cause missed detection risk or excessive detection cost. Therefore, the invention provides the AI-based product quality risk assessment method and system, which solve the problem that the traditional data driving model lacks physical interpretability by means of constructing a physical-guided cross-modal attention mechanism, constructing a manufacturing process risk propagation graph network, bayesian optimization self-adaptive threshold value and the like, realize the crossing from single-node alarm to risk chain retrospection, improve the accuracy and adaptability of risk assessment and provide a comprehensive, accurate and interpretable quality risk intelligent assessment solution for automobile bearing manufacturing enterprises. Disclosure of Invention The invention aims to provide an AI-based product quality risk assessment method and system. In order to achieve the above purpose, the invention is implemented according to the following technical scheme: The invention comprises the following steps: The method comprises the steps of obtaining multi-source data of an automobile bearing production chain, and carrying out deep extraction to obtain multi-source characteristics of the production chain, wherein the multi-source data of the production chain comprises raw material data, production line vibration signals, bearing structure parameters and visual images; Constructing a physical consistency penalty item according to the multi-source characteristics of the production chain, constructing a physically guided cross-modal attention matrix according to the multi-source characteristics of the production chain and the physical consistency penalty item, and carrying out modal fusion to obtain the state characteristics of the production chain; Extracting automobile bearing manufacturing process data, constructing a dynamic directed graph and a risk transmission adjacent matrix of the manufacturing process by combining production chain state characteristics, and carrying out risk transmission to determine risk values of all manufacturing processes, wherein the manufacturing processes comprise turning, heat treatment, grinding, assembly and detection; and carrying out Bayesian optimization on the risk threshold value of each manufacturing process according to the bearing structure parameters and the bearing type to obtain a risk dynamic threshold value, and carrying out quality risk assessment of the automobile bearing according to the risk value and the risk dynamic threshold value of each manufacturing process, wherein the quality risk assessment comprises overall risk assessment and manufacturing process risk assessment. Further, the method for obtaining the multi-source characteristics of the production chain by deep extraction comprises the following steps: the method comprises the steps of obtaining multi-source data of an automobile bearing production chain, wherein the multi-source data of the production chain comprises raw material data, production line vibration signals, bearing structure parameters and visual images; Comparing measured quality parameters of raw materials of each batch with corresponding standards, screening unqualified sampl