CN-121981611-A - Method for carrying out production quality root cause association analysis by utilizing IO-Link process data and final quality detection result

CN121981611ACN 121981611 ACN121981611 ACN 121981611ACN-121981611-A

Abstract

The application relates to the technical field of industrial automation and intelligent manufacturing, and discloses a production quality root cause association analysis method by utilizing IO-Link process data and a final quality detection result, aiming at solving the problem that the process data and the quality result are difficult to be associated accurately in the production process so as to position a quality root cause. The method comprises the steps of carrying out time slicing and feature extraction on IO-Link process data based on production beats, constructing a process feature vector sequence, aligning and labeling the process feature vector sequence with quality detection results of corresponding products, and generating a supervision learning sample set. According to the scheme, accurate fusion analysis of the process data and the quality results is realized, and the multivariate coupling anomalies can be identified from the high-dimensional data and the causal effects of the multivariate coupling anomalies are quantified, so that an interpretable and operable root cause analysis result is provided, and the efficiency and reliability of quality tracing and process optimization are improved.

Inventors

Liu Huaihao
LIU LIN
HAN XU
TONG YULIN
JI SIQI
ZHANG ZHENGYAN
Ge Manlin
HE JIE

Assignees

南京青尧自动化科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (6)

1. A production quality root cause association analysis method by utilizing IO-Link process data and a final quality detection result is characterized by comprising the following steps: Step S110, based on the production takt and the IO-Link data acquisition period, performing time slicing and feature extraction on multi-source heterogeneous IO-Link process data, and constructing a process feature vector sequence uniquely associated with a single product; Step S120, carrying out space-time alignment and labeling on the process feature vector sequence and a final quality detection result of a corresponding product to generate a supervised learning sample set with quality labels; Step S130, training the supervised learning sample set by adopting an integrated learning framework, and constructing a classification model capable of identifying a key feature subset strongly associated with the quality defect from the high-dimensional process features; Step S140, constructing a causal graph model from process feature anomalies to specific quality defect types based on the key feature subsets output by the classification model and combining a causal inference method, and quantifying the causal contribution degree of each process variable to the quality result; Step S150, according to the causal contribution degree and a preset contribution degree threshold, an interpretable root cause analysis report is generated, and a corresponding technological parameter adjustment instruction is triggered.
2. The method for performing production quality root cause association analysis by using IO-Link process data and final quality detection results according to claim 1, wherein in step S110, performing time slicing and feature extraction based on a product takt time and an IO-Link data acquisition period specifically includes: Firstly, determining a residence time window of each product at a specific station according to a start processing time stamp and an end processing time stamp of the product recorded by a production line PLC or an MES system; Secondly, taking the time window as a reference, intercepting all process variable data in a corresponding time period from original time sequence data streams acquired by all IO-Link master stations of the station; And finally, extracting statistical characteristics of time sequence data of each process variable in the time window, wherein the statistical characteristics comprise mean value, standard deviation, maximum value, minimum value, kurtosis, skewness and energy duty ratio of a specific frequency band extracted after fast Fourier transformation, and arranging all the extracted characteristics according to a preset sequence to form a multidimensional process characteristic vector.
3. The method for performing production quality root cause correlation analysis by using the IO-Link process data and the final quality detection result according to claim 1, wherein the step S110 further comprises performing dimension reduction and normalization on the extracted original feature vector; The dimension reduction treatment adopts a principal component analysis method, and principal components with accumulated contribution rate reaching more than 95% are reserved as new feature subsets; The normalization process adopts a Z-score normalization method to convert the numerical value of each feature into a distribution with the mean value of 0 and the standard deviation of 1, wherein the calculation formula is z= (x-mu)/sigma, wherein x is the original feature value, mu is the mean value of the feature in all samples, and sigma is the standard deviation of the feature in all samples.
4. The method for performing production quality root cause correlation analysis by using IO-Link process data and final quality detection results according to claim 1, wherein in step S130, the integrated learning framework is a gradient lifting decision tree; The training objective of the classification model is to minimize cross entropy loss between the predicted quality label and the real quality label; during model training, evaluating importance scores of each feature for classification tasks by calculating the sum of split gains of the feature in all decision trees; and after training is finished, screening out a key feature subset with importance scores higher than a preset importance score threshold.
5. The method for analyzing the root cause correlation of the production quality by using the IO-Link process data and the final quality detection result according to claim 1, wherein in the step S140, constructing a causal graph model by combining a causal inference method specifically includes: Taking the key feature subset screened in the step S130 as a candidate reason variable and the quality defect type as a result variable; analyzing the condition independence relation among variables in the supervised learning sample set by adopting a constraint-based causal discovery algorithm, and initially constructing a partial directed acyclic graph; Then, optimizing the primary graph structure by using a causal structure learning algorithm based on scores, and searching a causal graph structure which enables the Bayesian information criterion score to be optimal; In the final determined causal graph model, nodes represent process features or quality results, directed edges represent causal influence directions, and weights of the edges are estimated through a structural equation model or regression coefficients to obtain causal contribution degrees of all the process features to the quality results.
6. The method for analyzing production quality root cause correlation by utilizing IO-Link process data and final quality detection results according to claim 5, wherein the calculation of the causal contribution degree is based on a back-facts inference framework, and for a certain critical process feature Xi causing defects, the average causal effect is defined as an expected change amount of quality defect occurrence probability when Xi is interfered from an abnormal value Xiabnormal to a normal value Xinormal, and the expected change amount of quality defect occurrence probability is a causal contribution degree quantification index of the feature.

Description

Method for carrying out production quality root cause association analysis by utilizing IO-Link process data and final quality detection result Technical Field The invention relates to the technical field of industrial automation and intelligent manufacturing, in particular to a production quality root cause correlation analysis method by utilizing IO-Link process data and a final quality detection result. Background In modern intelligent manufacturing systems, quality control of the production process increasingly relies on deep mining of the inherent correlation between equipment operating conditions and product quality. The sensor and the actuator widely deployed in the industrial field collect massive process data through various communication protocols, wherein IO-Link is used as a standardized point-to-point serial communication interface, can provide abundant real-time process parameters including temperature, pressure, vibration, switch state and the like, and lays a data foundation for fine process monitoring. Meanwhile, the final quality detection link is usually completed by a visual system, a measuring instrument or manual spot check, so as to form a discrete but key product qualification judgment result. However, current manufacturing systems generally lack a technical means to effectively correlate high-frequency, fine-grained process data with low-frequency, result-oriented quality decisions, resulting in difficulty in quickly tracing root causes after quality problems occur, severely restricting the implementation of closed-loop quality improvement and intelligent decision-making capability. The correlation analysis of the production quality root cause by utilizing the IO-Link process data and the final quality detection result becomes a key technical direction for improving the transparency and the interpretability of the manufacturing process. The method aims at identifying key process characteristics and abnormal modes thereof which have obvious influence on the quality of products by carrying out structural modeling on time sequence process variables output by IO-Link equipment of each station on a production line and establishing a dynamic mapping relation with a downstream quality detection result. The prior art has multiple bottlenecks when the aim is achieved, namely firstly IO-Link process data has the characteristics of high dimension, isomerism and strong time sequence, quality detection results are mostly output in a label mode or a grading mode, natural mismatch exists between the data form and the time scale, an effective alignment and fusion mechanism is lacked, secondly, the traditional statistical process control method only focuses on single variable threshold value out-of-limit, hidden anomalies under multivariable coupling are difficult to capture, potential quality causes in complex processes cannot be revealed, and thirdly, the conventional root analysis is dependent on expert experience or static rule base, adaptability is poor when dynamic scenes such as product line configuration change, product model switching and the like are faced, and finally, most systems do not construct causal deducing paths from process anomalies to quality defects, so that analysis results stay in a correlation layer, and accurate process tuning and preventive intervention are difficult to support. These problems make it difficult for manufacturing enterprises to achieve efficient, reliable quality traceability and continuous optimization when dealing with flexible production demands of high frequency, small lot, and multiple varieties. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a production quality root cause correlation analysis method by utilizing IO-Link process data and a final quality detection result, and solves the problems in the background art. In order to achieve the above purpose, the invention provides the following technical scheme that the method comprises the following steps: Step S110, based on the production takt and the IO-Link data acquisition period, performing time slicing and feature extraction on multi-source heterogeneous IO-Link process data, and constructing a process feature vector sequence uniquely associated with a single product; Step S120, carrying out space-time alignment and labeling on the process feature vector sequence and a final quality detection result of a corresponding product to generate a supervised learning sample set with quality labels; Step S130, training the supervised learning sample set by adopting an integrated learning framework, and constructing a classification model capable of identifying a key feature subset strongly associated with the quality defect from the high-dimensional process features; Step S140, constructing a causal graph model from process feature anomalies to specific quality defect types based on the key feature subsets output by the classification model and combining a causal i