Reinforcing plate production risk perception and identification system based on big data analysis

By collecting full-dimensional production data and using weighted offset analysis and spatiotemporal graph convolutional networks, the limitations of multi-source data utilization and cross-dimensional feature association in reinforcement plate production are overcome, the accuracy of risk identification and prediction is improved, and the dynamic changes of the production process are adapted.

CN120822826AInactive Publication Date: 2025-10-21JIANGYIN SWIN ELECTRONICS NEW MATERIAL CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202510919682.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-04
Publication Date
2025-10-21
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing technologies lack the collaborative collection of multi-source data in reinforcement plate production and are unable to effectively identify potential connections between cross-dimensional factors, resulting in inaccurate risk feature screening and affecting production quality and safety.

Method used

Full-dimensional production data is collected through industrial Internet of Things terminals, and weighted offset analysis and self-service verification are used to achieve dynamic feature screening. The relationship between equipment topology and process chain is modeled using spatiotemporal graph convolutional networks, and the risk probability is dynamically corrected by combining conditional probability and association strength coefficient.

Benefits of technology

It achieves accurate identification of reinforcing plate production risks, reduces model complexity, improves prediction accuracy, adapts to dynamic changes in the production process, reduces detection costs and misjudgments, and supports refined risk differentiation and timely warning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120822826A_ABST
    Figure CN120822826A_ABST
Patent Text Reader

Abstract

The invention provides a stiffening plate production risk perception and identification system based on big data analysis, and relates to the technical field of stiffening plate risk perception, and the system comprises a data collection module which is used for an industrial Internet of Things terminal to collect and preprocess multi-source data in a stiffening plate production process; the feature extraction module is used for extracting features related to the production risk of the stiffening plate; the classification module is used for dividing equipment fault sub-classifications; the risk model construction module is used for training a stiffening plate production risk analysis model under the sub-classification through the first feature set under each sub-classification; the correlation model construction module is used for training a correlation analysis model of the stiffening plate production risk; the risk prediction module is used for inputting real-time data into each trained risk analysis model and obtaining a final risk probability according to the association analysis model; the early warning module is used for triggering the early warning module when the final risk probability exceeds an early warning threshold value; and an accurate risk identification solution is provided for intelligent manufacturing of the stiffening plate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of risk perception of reinforcement plates, and in particular to a reinforcement plate production risk perception and identification system based on big data analysis. Background Art

[0002] As a critical load-bearing component in industrial structures, reinforcing plates require multiple production processes, including stamping, welding, and heat treatment. These processes are susceptible to multiple factors, including equipment performance, raw material quality, process parameters, and environmental conditions, leading to quality defects such as cracks, deformation, and substandard hardness. With the advancement of smart manufacturing and Industrial Internet of Things (IIoT) technologies, reinforcing plate production has accumulated a vast amount of multi-source data (such as equipment vibration, process temperature, and raw material chemical composition). However, accurately identifying risk characteristics within this complex data and building cross-modal correlation models have become key challenges in improving production reliability.

[0003] Traditional systems collect only a single type of data (such as equipment vibration or temperature) and lack the coordinated collection of multiple sources of data, including raw material quality, operator behavior, and environmental parameters. This leads to a one-sided risk characterization. For example, it is impossible to correlate the combined effects of abnormal equipment speed and insufficient raw material toughness on crack defects.

[0004] Existing feature screening methods often use fixed thresholds (such as the Pearson correlation coefficient threshold) or static statistical methods (such as variance filtering). These methods fail to consider the dynamic temporal changes in production data (such as vibration feature drift caused by equipment aging). They only screen features within a single data dimension (such as equipment data or process data), severing the potential connections between cross-dimensional factors such as equipment failure, process fluctuations, and raw material problems. For example, they are unable to identify the synergistic impact of the combined features of "abnormal processing temperature + low raw material strength" on deformation defects.

[0005] How to screen out key features closely related to production risks from massive data has become the core link in building an accurate risk analysis model. Accurate feature screening can not only improve the prediction accuracy of the risk model, but also reduce computational complexity and improve system operation efficiency. It is of great significance to ensure the production quality of reinforcement plates, reduce production costs and improve production safety. Summary of the Invention

[0006] In response to the above problems, the present application proposes a reinforcing plate production risk perception and identification system based on big data analysis. It collects full-dimensional production data through industrial Internet of Things terminals, adopts weighted offset analysis and self-service verification to realize dynamic screening of features, uses spatiotemporal graph convolutional networks to model the relationship between equipment topology and process links, and realizes dynamic correction of risk probability through conditional probability and correlation strength coefficient, thereby breaking through the limitations of existing technologies in multi-source data utilization, cross-dimensional feature association and dynamic early warning, and providing accurate risk identification solutions for intelligent manufacturing of reinforcing plates.

[0007] The purpose of this application is achieved by the following technical solutions:

[0008] The present application provides a reinforcing plate production risk perception and identification system based on big data analysis, the system comprising:

[0009] A data acquisition module, which is used by the industrial Internet of Things terminal to collect and pre-process multi-source data from the reinforcing plate production process. The multi-source data includes production equipment operation data, raw material quality data, production process parameters, operator behavior data, production environment data, and quality inspection results;

[0010] A feature extraction module is used to extract features related to the production risk of reinforcing plates from the preprocessed data through feature engineering, wherein the features include time series features, statistical features, and frequency domain features;

[0011] The classification module is used to classify quality defects into subcategories based on quality inspection results and equipment failures into subcategories based on equipment data; under each subcategory, the corresponding parameter features are selected as the first feature set, and the first feature set is used as the input of the risk model under the subcategory;

[0012] A risk model building module is used to train a reinforcing plate production risk analysis model under each sub-category using the first feature set under the sub-category;

[0013] A correlation model building module, used for training the correlation analysis model of the reinforcing plate production risk;

[0014] The risk prediction module is used to input the real-time collected and pre-processed production data into the trained risk analysis models to obtain preliminary risk probabilities. The preliminary risk probabilities are adjusted based on the output of the association analysis model, and the adjusted quality defect probability or equipment failure probability of the current production process is output as the final risk probability.

[0015] The early warning module is used to trigger the early warning module when the final risk probability exceeds the early warning threshold.

[0016] The beneficial effects of the present invention include at least: covering six dimensions of data: equipment operation, raw materials, process parameters, personnel behavior, environment and quality inspection, and constructing a unified time series data set through preprocessing such as time-space alignment and missing value filling, thereby avoiding analytical deviations in a single data dimension (such as only focusing on equipment temperature and ignoring the impact of raw material toughness on cracks).

[0017] Through a two-dimensional analysis of "mean shift + standard deviation shift," combined with a dynamic threshold (a first preset proportional coefficient adjusted with sample size) and bootstrap verification, features strongly associated with risk (such as the contribution of stamping pressure fluctuations to dimensional deviation) are screened out, reducing model complexity while improving prediction accuracy. A conditional probability matrix is ​​constructed based on historical co-occurrence data to quantify the bidirectional association between defects and failures. This is combined with a spatiotemporal graph convolutional network to model equipment topology and process links (such as the material transfer delay between a stamping machine and a welding robot), capturing the multi-level risk transmission path of "equipment failure → process anomaly → quality defect." The correlation strength coefficient is dynamically output to adapt to process adjustments or equipment aging (for example, the correlation strength between mold wear and edge cracks is automatically updated after a new mold is put into use). Through the coupled calculation of the correlation strength coefficient and conditional probability, the risk of the linkage between equipment failure and quality defects is quantified, avoiding misjudgments from a single model. Risk probability and correlation strength are combined to divide warning levels, matching different response strategies to achieve a refined differentiation of risk urgency. The basic warning threshold is dynamically adjusted based on historical defect frequency (the higher the frequency, the lower the threshold), adapting to both the initial commissioning phase of the production line (lowering the threshold for high-frequency defects, providing early warning) and the stable operation phase (raising the threshold for low-frequency defects, reducing false alarms). Inspection resources are allocated based on the adjusted risk probability (for example, increasing the frequency of online inspections for high-risk defects), reducing inspection costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of a reinforcement plate production risk perception and identification system based on big data analysis provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] Below, the present application is further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0020] See also Figure 1 The present application provides a reinforcing plate production risk perception and identification system based on big data analysis, the system comprising:

[0021] A data acquisition module, which is used by the industrial Internet of Things terminal to collect and pre-process multi-source data from the reinforcing plate production process. The multi-source data includes production equipment operation data, raw material quality data, production process parameters, operator behavior data, production environment data, and quality inspection results;

[0022] A feature extraction module is used to extract features related to the production risk of reinforcing plates from the preprocessed data through feature engineering, wherein the features include time series features, statistical features, and frequency domain features;

[0023] The classification module is used to classify quality defects into subcategories based on quality inspection results and equipment failures into subcategories based on equipment data; under each subcategory, the corresponding parameter features are selected as the first feature set, and the first feature set is used as the input of the risk model under the subcategory;

[0024] A risk model building module is used to train a reinforcing plate production risk analysis model under each sub-category using the first feature set under the sub-category;

[0025] A correlation model building module, used for training the correlation analysis model of the reinforcing plate production risk;

[0026] The risk prediction module is used to input the real-time collected and pre-processed production data into the trained risk analysis models to obtain preliminary risk probabilities. The preliminary risk probabilities are adjusted based on the output of the association analysis model, and the adjusted quality defect probability or equipment failure probability of the current production process is output as the final risk probability.

[0027] The early warning module is used to trigger the early warning module when the final risk probability exceeds the early warning threshold.

[0028] In one possible implementation, the production equipment operation data includes equipment speed, temperature, vibration frequency, and operation time; the raw material quality data includes the strength, toughness, and chemical composition of the raw materials; the production process parameter data includes processing temperature, pressure, time, and processing speed; the operator behavior data includes operation step compliance, operation time, and operation frequency; the production environment data includes ambient temperature, humidity, and dust concentration; and the quality inspection results include defect type, location, and size.

[0029] The preprocessing includes performing spatiotemporal alignment on the collected multi-source heterogeneous data, and constructing a unified format of the production process time series data set through missing value filling, outlier removal, noise filtering and standardization;

[0030] Among them, the Laida criterion, interquartile range method or isolation forest algorithm can be used to identify and remove abnormal data;

[0031] Use mean imputation, median imputation, multiple imputation, or machine learning-based imputation methods to handle missing data;

[0032] The data were normalized using the minimum-maximum normalization, Z-score normalization, or decimal scaling normalization methods.

[0033] The working principle of the above technical solution is:

[0034] Real-time data collection of the entire reinforcing plate production process is carried out through industrial Internet of Things terminals, covering six major dimensions: equipment, raw materials, processes, personnel, environment, and quality inspection. The sampling timestamps and spatial locations of different devices and sensors are unified to ensure data temporal consistency (such as aligning the sampling data of each device every minute to the same time point). The Laida criterion (3σ principle) is used to identify data that exceeds ±3 times the standard deviation of the mean. For example, when the equipment temperature is continuously higher than the normal range by 3°C, it is judged as abnormal and eliminated. For missing values ​​of raw material strength data, the median interpolation method is used (taking the median of the strength of the raw materials in the batch to fill in the missing values), and the processing speed data is standardized, such as Z-score standardization, to make data of different dimensions comparable.

[0035] Extract trends (e.g., increasing speed for five consecutive hours) and periodicity (e.g., hourly fluctuations in vibration frequency) from equipment operating time data. Calculate the mean, variance, and quantiles of raw material strength (e.g., the 95th percentile indicates a high-strength threshold). Perform a Fourier transform on the equipment vibration signal to extract the proportion of high-frequency components (e.g., an increase in the proportion of frequency components above 100 Hz may indicate bearing wear).

[0036] Based on the quality inspection results, quality defects are divided into subcategories such as cracks, deformation, and dimensional tolerance. Based on the equipment data, faults are divided into subcategories such as bearing wear and motor overheating. For different subcategories, different parameters are selected as the first feature set;

[0037] The risk analysis model training module uses algorithms such as random forest and LSTM, taking the first feature set as input and quality inspection results or equipment failure records as labels to train predictive models for each subcategory. For example, the model is trained using historical crack defect data, inputting features such as processing temperature and pressure, and outputting the crack risk probability. Association analysis model training utilizes the Apriori algorithm or graph neural network to explore the correlations between different factors. For example, it was found that when the raw material toughness is less than 80 MPa and the processing pressure is greater than 100 MPa, the probability of crack occurrence increases by 40%, and this association rule is incorporated into the model.

[0038] Real-time data is fed into each sub-classification risk model to generate a preliminary probability (e.g., a crack risk of 60%). This preliminary probability is then adjusted based on the factor correlation results output by the correlation model. For example, if the preliminary crack risk prediction is 60%, and the correlation model indicates a high risk correlation between the current processing temperature and the raw material toughness, the probability is adjusted to 80%. A warning threshold (e.g., 70%) is set, and when the final risk probability exceeds the threshold, an alert is triggered via an audible or visual alarm or system notification.

[0039] By integrating cross-dimensional data from equipment, raw materials, and processes, analytical biases from a single data dimension are avoided (e.g., simply looking at equipment temperature fails to reveal the impact of raw material quality on defects). Classification modules are used to divide data into subcategories, and specialized models are trained based on subcategory characteristics, reducing data volume while improving prediction accuracy (e.g., different defect types have different influencing factors and require specialized models). Correlation models are combined to uncover hidden relationships (e.g., the indirect correlation between human operations and equipment failures), making risk predictions more relevant to actual production scenarios. Through data-driven algorithmic modeling, the system transforms hidden risks in industrial production into quantifiable probabilistic indicators, enabling a shift from post-detection to pre-warning. Typical applications include stamping die failure warnings in automotive reinforcement plate production and crack risk control in aviation material reinforcement plates.

[0040] In a possible implementation, the classification module implementation steps include:

[0041] The first sub-category is divided into the types of defects; including cracks, deformation, substandard hardness and dimensional deviation;

[0042] The second sub-category is divided according to the type of equipment anomaly; including mechanical failure, electrical failure, and sensor failure;

[0043] Under each sub-category, the mean and standard deviation of multiple defective or abnormal samples of each parameter feature are calculated as the first mean and first standard deviation;

[0044] The mean and standard deviation of each parameter characteristic when the quality inspection is passed and the equipment is operating normally are used as the second mean and second standard deviation;

[0045] A first feature set of a sub-category is determined according to a first mean, a first standard deviation, a second mean, and a second standard deviation of a plurality of parameter features under the sub-category.

[0046] In a possible implementation, determining the first feature set of a sub-category according to a first mean, a first standard deviation, a second mean, and a second standard deviation of a plurality of parameter features under a sub-category includes:

[0047] In each sub-category, the deviation ratio of the first mean value of each parameter relative to the second mean value is calculated as the first deviation value; first deviation value = |first mean value - second mean value | / second mean value;

[0048] The offset ratio between the first standard deviation and the second standard deviation of the parameter is used as the second offset; the second offset = |first standard deviation - second standard deviation| / second standard deviation;

[0049] Sort the multiple first offsets from high to low, and calculate the sum of the multiple first offsets;

[0050] Calculate the sum of all sorted first offsets and multiply this sum by a first preset proportional coefficient to obtain a first threshold. The first preset proportional coefficient can decrease as the number of defect samples increases. The first preset proportional coefficient can be 80% (for example, first preset proportional coefficient = 80% - (sample size - 500) / 1000 * 10%). The use of an 80% first preset proportional coefficient stems from the Pareto optimality theory (80 / 20 rule), which states that a small number of key features (approximately 20%) explain most of the risk variation (approximately 80%). By screening features using the cumulative contribution threshold, model complexity can be reduced while retaining explanatory power.

[0051] Select the first H parameter features from the sorted first offsets, so that the sum of the first offsets of the H parameter features reaches or exceeds the threshold for the first time, and use the first H parameter features as the first candidate features of the sub-classification;

[0052] The first candidate feature and the second offset are combined to determine a final first feature set.

[0053] In a possible implementation, combining the first candidate feature and the second offset to determine the final first feature includes:

[0054] The sum of the second offsets of the parameters in the first candidate feature is calculated and multiplied by a second preset proportional coefficient to obtain a second threshold value. The second preset proportional coefficient may be 50%. The second offset reflects the impact of parameter fluctuations on risk. Taking 50% as the threshold value can balance the ability to identify "volatility-sensitive risk" and "mean-shift risk."

[0055] If the second deviation of a parameter exceeds the mean and its volatility contribution exceeds 50% of the total volatility, then the parameter has a significant impact on the instability of the risk;

[0056] Set the first offset threshold coefficient α, and calculate α times the first threshold as the third threshold; where 0<α<1,

[0057] Remove parameters from the first candidate feature set that simultaneously satisfy the first offset less than the third threshold and the second standard deviation greater than the second threshold; exclude parameters with large fluctuations but small mean shift (probably just noise); retain parameters with large mean shift and stable fluctuations (clear risk signals);

[0058] The remaining parameter features constitute the final first feature set;

[0059] When the candidate features do not include the key process parameters, the key process parameters are forcibly added to the first feature set; wherein the key process parameters are determined based on the degree of correlation between each process parameter and the quality defect equipment failure and expert experience.

[0060] In a possible implementation, the system further performs the following steps:

[0061] Perform correlation analysis on the first candidate feature set and remove redundant features;

[0062] Bootstrap method was used to verify the stability of feature screening results;

[0063] The first proportional coefficient, the second proportional coefficient, and the threshold coefficient are updated periodically according to new data.

[0064] The working principle and effects of the above technical solution are as follows:

[0065] The dual-dimensional offset analysis accurately distinguishes risk signals from noise through cross-validation of mean shift and volatility stability. Its core logic is to locate significant abnormal parameters using mean shift and then remove high volatility interference items using standard deviation shift, ensuring that the final feature set only contains parameters with strong risk associations. The specific working principle is as follows:

[0066] First offset: anchoring significant anomaly parameters

[0067] Prioritize calculating the mean difference between defective and qualified samples for each parameter (first offset = |first mean - second mean | / second mean) and sorting them in descending order by offset. This step, centered on steady-state deviation, quickly identifies parameters directly associated with risk. For example, if the first mean of the punching temperature of a reinforcement plate is consistently 15% higher than the mean of qualified samples, it is highly likely to cause material cracking or deformation, and such parameters will be prioritized for inclusion in the first candidate feature set.

[0068] Second offset: filter high fluctuation noise

[0069] For the parameters in the first candidate feature set, calculate their standard deviation offset (second offset = |first standard deviation - second standard deviation| / second standard deviation) and filter them based on the second threshold. This step focuses on identifying interference items with abnormal fluctuations but small impact on the mean:

[0070] If a parameter's second offset is extremely high (violent fluctuations) but its first offset is below the third threshold (little impact on the mean), it is considered random noise. For example, fluctuations in ambient temperature and humidity may cause a sharp increase in the standard deviation of sensor data, but have no substantial impact on product quality. Such parameters will be excluded.

[0071] If a parameter meets both the first high deviation (significantly abnormal mean) and the second low deviation (stable fluctuation), it is confirmed as a strong risk signal. For example, if the mean of the welding current is consistently 20% higher than the standard value and the standard deviation does not change significantly, it indicates that this parameter is steadily deviating from the normal range and is the core factor causing welding defects, and will be retained in the final feature set.

[0072] The mean shift priority screening mechanism ensures that parameters that directly cause defects are not missed; the fluctuation stability check eliminates random interference and reduces the model misjudgment rate.

[0073] For key parameters pre-defined by process experts (such as stamping pressure and heating temperature), the system is forced to retain them even if their double offset does not reach the threshold. This mechanism ensures that core process parameters are not accidentally deleted due to data fluctuations, further enhancing the reliability of risk identification. Through the dual verification of mean shift and standard deviation shift, the system achieves progressive screening from "rough screening of anomalies" to "precise confirmation numbers", significantly improving the purity and predictive ability of the risk feature set.

[0074] When the sample size is small (e.g., <500), a higher coefficient (80%) is used to ensure that enough features are captured to avoid information loss due to data sparsity. As the sample size increases (e.g., >1500), the coefficient is reduced to 70%, allowing for a more streamlined feature set, reducing model complexity and preventing overfitting.

[0075] The second preset proportional coefficient (50%) balances the fluctuation sensitivity, ensuring that the system's response to parameter fluctuations is neither too sensitive (to avoid false positives) nor too slow (to avoid false negatives), thereby optimizing the risk warning score.

[0076] Key process parameters (such as temperature and pressure) are pre-determined based on domain knowledge and historical data analysis. Even if these parameters are not selected in data-driven screening, they will be forced to be added to the feature set. For example, in the stamping process of the reinforcement plate, if the "stamping pressure" is not selected as a high-offset feature due to data fluctuations, the forced addition mechanism can still ensure that this parameter is involved in risk modeling, avoiding prediction failure due to the missing core process parameters.

[0077] By filtering features using the Pareto principle (80 / 20), the original high-dimensional parameter space is compressed to a subset of key features, significantly reducing the training and inference costs of subsequent risk models. Parameters that simultaneously meet the "low mean shift" and "high volatility" criteria are removed. These parameters are often random noise or interference factors that are weakly correlated with risk.

[0078] Multiple sampling validations ensure that the selected feature set remains stable across different data subsets, reducing randomness in feature selection due to data bias and improving model generalization. The selected feature set is highly consistent with physical process logic (e.g., temperature affects material hardness, and pressure affects deformation), consistent with engineers' understanding of the production process, and making risk prediction results more interpretable. When the system issues an alert, engineers can quickly locate the specific process parameter anomaly, shortening troubleshooting time (e.g., from hours to minutes).

[0079] In a possible implementation, the first mean is a weighted mean of values ​​of multiple defective or abnormal samples, where the weights are determined by:

[0080] The time weight is calculated based on the test time of the defect sample or abnormal sample. The closer the test time is to the current time, the higher the time weight is.

[0081] Calculating a grade weight based on the severity of the defect or anomaly, wherein the higher the severity, the higher the grade weight; wherein the grade weight is determined by the proportion of the defect size exceeding a threshold;

[0082] For example: Calculate the defect size exceeding the standard ratio CB v =max(0,(Ds v -Dst hr ) / Dst hr ); The excess ratio is mapped to the grade weight through the Sigmoid function; where Ds v is the size of the vth defect sample; Dst hr is the size threshold;

[0083] The final weight of the sample data is determined based on the time weight and level weight.

[0084] For example, the final weight may be the product of the time weight and the level weight, or the final weight may be the weighted mean of the time weight and the level weight.

[0085] Time weighting ensures the model is more sensitive to recent defects and adapts to the timeliness of equipment performance degradation. For example, the increase in defect frequency caused by mold wear can be captured promptly. Grade weighting causes the contribution of severe defects to the mean to grow nonlinearly, preventing the risk of "small defects with a large number of minor impacts" from masking the true impact of "large defects with a large impact." Lower weighting is assigned to early minor defects to reduce the interference of historical noise data on current risk assessments and enhance the model's noise tolerance. This weighted mean algorithm, by integrating dynamic weights for time and severity, transforms traditional static statistics into dynamic assessments of the current risk status, providing more accurate basic data support for the reinforcing plate production risk perception and identification system.

[0086] Risks are broken down into multiple subcategories, including quality defects (cracks, deformation, etc.) and equipment failures (mechanical / electrical failures, etc.), with independent model training for each subcategory. This design accurately captures the unique data patterns of different risk types and avoids information loss caused by one-size-fits-all modeling.

[0087] By calculating the mean offset (first offset) and standard deviation offset (second offset) of parameter features, and combining them with a dynamic threshold (a first preset proportional coefficient that changes with the number of defect samples), key features are screened. This mechanism can adaptively adjust the feature importance distribution of different sub-categories, ensuring that the features input to the model are highly correlated with the risks.

[0088] Time decay weights (higher weights for recent samples to be tested) and severity weights (determined by the proportion of defect size exceeding the threshold) are used for defect / abnormal samples, so that model training focuses more on recent and serious risk events; this design enhances the model's ability to respond to real-time changes in the production system.

[0089] In one possible implementation, the steps of implementing the risk model building module include:

[0090] For each subclassification, the first feature set dataset is divided into a training set and a validation set according to a preset ratio when the number of defect / fault samples is greater than or equal to the sample size threshold. When the number of defect / fault samples is less than the sample size threshold, a combination of synthetic minority class oversampling technique (SMOTE) and random undersampling is used. Minority class samples are first generated using SMOTE with feature constraints. The constraints include: the feature values ​​of the generated samples must meet the physical boundaries of the process (for example, the temperature feature must be within the rated operating range of the equipment); and the feature correlation of the generated samples must be consistent with that of the original samples.

[0091] For quality defect subcategories such as cracks and deformations, a time-sensitive LSTM neural network is used and an attention mechanism is added;

[0092] For equipment failure subclasses, the XGBoost ensemble learning algorithm is used;

[0093] The final weight W of each feature parameter is obtained by the first offset, the second offset, the number of defect samples associated with the feature, and the time decay coefficient in the first feature set. final ; The final weight is used to adjust the input weight of each feature in the training data;

[0094]

[0095] Among them, δ u is the first offset of a characteristic parameter, δ σ is the second offset of the feature parameter; ε is a constant, 0<ε<0.02, ε is a regularization term to prevent the denominator from being zero, YN is the number of defect samples associated with the feature; η model is the feature importance score output by the selected algorithm; λ is the time decay coefficient, 0<λ≤0.1; wmin is the minimum weight threshold of historical data; Δtk is the difference between the center time of time window k and the current time; YN kis the number of defect samples associated with the feature within time window k; K is the total number of time windows. The windows can be divided into exponential windows (such as 7 days, 30 days, and 90 days), which are suitable for attenuation-sensitive scenarios, or fixed windows (such as weekly and monthly) to facilitate business understanding and analysis. Bayesian online change point detection (BOCD) can also be used to identify data distribution mutation points in the time series of defect samples associated with the feature. Time windows are automatically divided based on the mutation points.

[0096] Different algorithms calculate η through different mechanisms model ,For example:

[0097] Tree models (XGBoost) based on split gain or frequency;

[0098] The attention model (LSTM+Attention) is based on weight distribution;

[0099] The general method (SHAP) is based on game-theoretic attribution.

[0100] In the dynamic weighted training step, the time decay coefficient λ is dynamically adjusted according to the stability of the production process. The more frequent the process changes, the larger the λ value;

[0101] The Bayesian optimization algorithm is used to search for the optimal hyperparameter space, and the distribution differences between newly collected data and historical data are regularly detected. For example, when the KS test p-value is <0.01, incremental model training is triggered.

[0102] The effects of the above technical solution are:

[0103] Flexible processing methods are selected based on the number of defect / fault samples. When the sample size is sufficient, training and validation sets are routinely divided to ensure the foundation of model training. When the sample size is insufficient, a combination of SMOTE with feature constraints and random undersampling is used to solve the problem of few samples. At the same time, through process physical boundaries and feature correlation constraints, the generated samples are made to fit the actual production logic, maintain data authenticity, and improve the model's ability to identify niche risks.

[0104] SMOTE under feature constraints prevents the generation of invalid samples that are out of touch with actual production, making data enhancement in small-sample scenarios more scientific, providing high-quality data for subsequent model training, and reducing model bias caused by unreasonable samples.

[0105] For quality defect subcategories such as cracks and deformations, the LSTM neural network combined with the attention mechanism is in line with their temporal change characteristics and can effectively capture the risk patterns that evolve over time in the production process. The attention mechanism can also focus on key time series segments and strengthen the identification of key risk nodes; the equipment failure subcategory uses XGBoost integrated learning, taking advantage of its efficient processing of complex nonlinear relationships and anti-overfitting capabilities, accurately mining the correlation between equipment failure-related features, and improving the accuracy of fault prediction.

[0106] Different risk subcategories are adapted with exclusive algorithms to give full play to the advantages of the algorithms. Compared with a single algorithm, it can cover more diverse risk models and enhance the overall risk model's ability to identify various risks in the reinforcement plate production process.

[0107] The final weight is calculated by integrating the first offset, second offset, number of associated defect samples, time decay coefficient, etc., and the relationship between features and risks is considered in multiple dimensions; the first and second offsets reflect the degree of deviation between the feature mean and standard deviation, highlighting the abnormality of the feature; the number of associated defect samples reflects the close relationship between the feature and the risk; the time decay coefficient adapts to dynamic changes in production, making the impact of recent data more prominent, and comprehensively and accurately measuring the role of features in risk identification, optimizing the input weight of training data, and making the model more focused on key risk features.

[0108] The time decay coefficient is adjusted according to the stability of the production process, and the decay is increased when the process changes frequently, so as to timely track the impact of process changes on feature importance. The feature importance scores output by different algorithms are combined, and the advantages of multiple algorithms are compatible, so that the weight calculation is more comprehensive and dynamic, so that the model can continuously adapt to changes in the production process and ensure the accuracy and timeliness of risk identification.

[0109] The Bayesian optimization algorithm searches for the optimal hyperparameter space to improve model training efficiency and effectiveness; data distribution difference detection is regularly performed, such as the KS test triggering incremental training, so that the model can adapt to new data in a timely manner, cope with data drift in the production process, maintain good risk identification performance, and extend the effective life cycle of the model.

[0110] From sample processing and algorithm selection to weight setting and model iteration, a complete and dynamic optimization system is built to enable the risk model to stably and accurately identify risks in the complex and ever-changing reinforcement plate production environment, provide reliable support for production decisions, and help companies prevent and control quality defects and equipment failures in advance, thereby reducing production losses.

[0111] In a possible implementation, the association model building module includes:

[0112] Construct defect-fault co-occurrence labels based on historical data; calculate the conditional probabilities of different defect types and fault types; construct a defect-fault coupling matrix; the conditional probabilities include: P(defect|fault) and P(fault|defect);

[0113] P(defect|fault) is P(D i 丨F j ): the probability of a specific defect occurring when a certain type of fault occurs in the device, and the calculation formula is:

[0114]

[0115] P(fault|defect) is P(F j 丨D i ): the probability of the corresponding device fault when a certain type of defect exists in the product, and the calculation formula is:

[0116]

[0117] where F j is the j-th type of device fault, and D i is the i-th type of quality defect;

[0118] Construct a defect-fault coupling matrix with the conditional probabilities as elements:

[0119] [[ID=三十五]]

[0120] Use a spatio-temporal graph convolutional network to model the device topology and process link, and output the cross-modal correlation strength coefficient β ij ∈[0,1].

[0121] β ij The larger it is, the stronger the correlation;

[0122] The construction of the spatio-temporal graph convolutional network includes:

[0123] Nodes: production equipment units + quality inspection workstations;

[0124] Edge weights: material transfer delay + signal correlation;

[0125] Dynamic features: device real-time state encoding + environmental fluctuation amount.

[0126] Combine cross-entropy loss and reconstruction loss to optimize the model parameters;

[0127] where the loss function is: Loss = y·LCE+(1 - y)·Lrecon

[0128] LCE measures the difference between the predicted correlation strength and the actual co-occurrence probability; Lrecon ensures that the model can reconstruct the original input features, and y is the weight coefficient.

[0129] The working principle and effects of the above technical solution are as follows:

[0130] Based on historical production data, the simultaneous occurrence of equipment failure events and quality defect events is marked to form a "defect-failure" co-occurrence label, providing basic data for subsequent probability calculations. Explanation of the conditional probability formula: P(D i 丨F j ): Taking equipment failure as a premise, calculate the probability of a specific defect occurring. For example, if the "bearing wear" failure occurs 100 times in historical data, and among them, the "crack" defect appears 20 times, then P(D i 丨F j ) = 20 / 100 = 0.2, indicating that the probability of a crack occurring when the bearing is worn is 20%. P(F j 丨D i ): Taking quality defect as a premise, calculate the probability of the corresponding equipment failure. For example, if the "deformation" defect appears 80 times, and among them, the "motor overheating" failure occurs 16 times, then P(F j 丨D i ) = 16 / 80 = 0.2, indicating that the probability of motor overheating when the deformation defect appears is 20%. Combine all conditional probabilities into a two-dimensional matrix to visually present the two-way probability relationship between defects and failures. For example, the element in the first row and first column of the matrix reflects the probability that the "bearing wear" failure leads to the "crack" defect.

[0131] Abstract production equipment (such as stamping machines, welding robots) and quality inspection stations as graph nodes, and each node contains the real-time state of the equipment (such as temperature, pressure) and historical failure records.

[0132] Based on the material transfer path and process sequence, construct directed edges between nodes, and the edge weight is the material transfer time delay (such as the transfer time from the heating furnace to the stamping machine); calculate the mutual information between equipment parameters and defect types (such as the mutual information value between stamping pressure fluctuation and sheet crack) through historical data as the logical weight of the edge. Feature extraction and propagation of node features Encoding: Static features include equipment type, process parameters, thresholds, etc.; dynamic features include real-time sensor data, fault alarm records, etc.; aggregate adjacent node information to capture the synergy between equipment (such as the impact of upstream equipment failure on downstream quality); process time series data through a gated recurrent unit (GRU) to extract the time-dependent relationship between failures and defects (such as how long after a failure a defect appears); the model learns the contribution degree of each node to the target defect through a self-attention mechanism to generate association weights; map the output of the spatio-temporal graph convolution to the [0,1] interval through the Sigmoid function.

[0133] By combining the defect-fault coupling matrix with the spatiotemporal graph convolutional network, the system can capture the complex cause-and-effect chain in the production process (such as "sensor failure → parameter misjudgment → temperature abnormality → hardness substandard") and predict multi-level risk transmission in advance.

[0134] Conditional probabilities are presented in matrix form, visually demonstrating the many-to-many relationships between defects and faults, supporting the rapid location of risk transmission paths. Nodes cover production equipment units and quality inspection stations, and edge weights integrate material transmission delays and signal correlations to depict the physical and logical connections within the production chain.

[0135] Combine the real-time status of equipment (such as temperature and vibration) with environmental fluctuations (such as humidity and dust) to capture the risk evolution pattern in the time dimension.

[0136] The spatiotemporal graph convolutional network (ST-GCN) learns the spatiotemporal coupling relationship between defects and faults, quantifies the correlation strength between different device-defect pairs, and supports dynamic risk tracing. The coupling matrix complements the output of the ST-GCN: the matrix provides static statistical correlation, while the ST-GCN provides dynamic temporal correlation, comprehensively improving the accuracy of risk prediction.

[0137] When a device fails, the real-time β ij When the threshold is exceeded, the defect warning probability is dynamically adjusted in combination with P(defect|fault) to avoid misjudgment by a single model. For example, if a mechanical failure occurs, ST-GCN outputs β(crack|mechanical failure) = 0.8, and the historical conditional probability P(crack|mechanical failure) = 0.75, then the warning probability of the crack defect is significantly improved.

[0138] By combining the coupling matrix with ST-GCN analysis, high-risk correlation paths (e.g., mechanical vibration → curing temperature fluctuation → crack defects) can be identified, enabling targeted adjustments to process parameters. By combining real-time environmental and equipment status, potential correlation risks (e.g., dimensional deviation caused by sensor failure in high-temperature and high-humidity environments) can be predicted, enabling proactive intervention in the production process.

[0139] In one possible implementation, the steps of implementing the risk prediction module include:

[0140] Input the real-time collected and pre-processed production data into the trained risk analysis models to obtain the preliminary risk probability P ini (D i ) and P ini (F j );

[0141] Adjust the preliminary risk probability based on the conditional probability and correlation strength coefficient output by the correlation analysis model;

[0142] When a device failure is detected F jThe preliminary risk probability is P ini (F j ) when the corresponding quality defect D i The risk probability is adjusted to P adj (D i ):

[0143] P adj (D<( i ) = P ini (D i ) + β ij ·P ini (F j )·P(D i 丨F j )

[0144] When the preliminary risk probability of the detected quality defect D i is P ini (D i ), the risk probability of the corresponding equipment failure F j is adjusted to P adj (F j ):

[0145] P adj (F j ) = P ini (F j ) + β ij ·P ini (D i )·P(F j 丨D i [[ID=6,7]])

[0146] Output the adjusted probability of quality defect or equipment failure in the current production process as the final risk probability;

[0147] [[ID=,73]]Allocate inspection resources according to the final risk probability.

[0148] The effect of the above technical solution is: by mining the implicit association between the equipment and the defect through the β ij coefficient. For example, when the preliminary probability of the equipment "abnormal motor temperature" is 0.3, and the β ij with "weld joint void" is 0.8, P(D i 丨F j ) = 0.6, the risk of D i is increased from 0.2 to P adj = 0.2 + 0.8×0.3×0.6 = 0.344, and potential quality problems are discovered in advance.

[0149] When equipment parameters are abnormal, it not only warns of equipment failure but also predicts the chain of quality defects that may arise. For example, "hydraulic system pressure fluctuations" trigger an "uneven plate thickness" warning through associated paths, avoiding losses in downstream processes. When a quality defect is detected, it reversely infers the most likely source of the equipment failure.

[0150] β ij The coefficient is updated with production data to adapt to process adjustments or equipment aging. For example, when a new mold is put into use, the correlation strength β between "mold wear" and "edge cracks" is ij From 0.5 to 0.7, the system automatically increases the sensitivity of relevant warnings.

[0151] Conditional probability filters random fluctuations to avoid false positives. For example, short-term fluctuations in ambient temperature can cause sensor data anomalies, but because P(defect | temperature fluctuation) is extremely low, the ultimate risk probability is not significantly affected. Risk propagation path visualization uses formulas to analyze risk transmission paths, helping engineers understand cause-and-effect relationships. By clearly demonstrating the risk transmission mechanism from "abnormal vibration to cracks," it assists in developing targeted prevention and control measures.

[0152] Allocate testing resources based on the adjusted risk probability; for example, adj (D i )>0.8 for key defects, increase the frequency of online inspection, and adj (F j )<0.3, the inspection cycle will be extended for minor faults.

[0153] Schedule maintenance in advance by analyzing the probability of equipment failure and quality defects.

[0154] In one possible implementation, the steps for implementing the early warning module include:

[0155] When the final risk probability exceeds the basic warning threshold, the warning module is triggered; the warning level is determined based on the risk probability and the correlation strength coefficient;

[0156] For example,

[0157] If the final risk probability is greater than or equal to 0.7 and the correlation strength coefficient is greater than or equal to 0.6, a red alert is triggered and the system is immediately shut down for investigation;

[0158] If the final risk probability is greater than or equal to 0.5 and less than 0.7, and the correlation strength coefficient is greater than or equal to 0.4, an orange warning is triggered and manual review is required within the preset time;

[0159] If the final risk probability is greater than or equal to 0.3 and less than 0.5, and the correlation strength coefficient is less than 0.4, a yellow warning is triggered, the risk is marked, and continuous monitoring is required;

[0160] Among them, the basic warning threshold is dynamically adjusted according to the historical frequency of defects.

[0161] The higher the frequency of defects, the more serious the level, and the lower the basic warning threshold.

[0162] The basic warning threshold is dynamically adjusted based on historical defect frequency, with higher defect frequencies and more severe levels resulting in lower thresholds. This design closely aligns with the actual risk profile of the production line, avoiding false alarms or missed alerts caused by fixed thresholds. During initial production line commissioning or when process fluctuations are significant, high-frequency, high-severity defects will cause the threshold to drop rapidly, triggering early warnings and helping companies identify potential risks. During stable production line operation, lower defect frequencies will maintain higher thresholds, reducing unnecessary warning interference and ensuring production efficiency.

[0163] The warning level is determined by combining the final risk probability and the correlation strength coefficient, achieving a refined assessment of the risk. The risk probability reflects the possibility of the current risk occurring, while the correlation strength coefficient reflects the degree of mutual influence between risks. When both are at a high level (such as the red warning condition), the system determines that the risk is serious and highly correlated, and immediately triggers a shutdown for investigation to minimize the losses caused by the risk; orange and yellow warnings are based on different threshold combinations, and graded measures such as manual review and continuous monitoring are adopted to ensure that enterprises can reasonably allocate resources according to the degree of risk and strike a balance between risk prevention and control and production costs.

[0164] The working principle and effect of the above technical solution are the same as those in the embodiment of the method of this application, and will not be repeated here.

[0165] This application is explained from the perspectives of purpose of use, effectiveness, progress and novelty, and has complied with the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings of this application are only preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc. that are similar or identical to those of this application, that is, all equivalent replacements or modifications made in accordance with the scope of the patent application of this application, should fall within the scope of protection of the patent application of this application.

Claims

1. A reinforcement plate production risk perception and identification system based on big data analysis, characterized in that: The system comprises: A data acquisition module, which is used by the industrial Internet of Things terminal to collect and pre-process multi-source data from the reinforcing plate production process. The multi-source data includes production equipment operation data, raw material quality data, production process parameters, operator behavior data, production environment data, and quality inspection results; A feature extraction module is used to extract features related to the production risk of reinforcing plates from the preprocessed data through feature engineering, wherein the features include time series features, statistical features, and frequency domain features; The classification module is used to classify quality defects into subcategories based on quality inspection results and equipment failures into subcategories based on equipment data; under each subcategory, the corresponding parameter features are selected as the first feature set, and the first feature set is used as the input of the risk model under the subcategory; A risk model building module is used to train a reinforcing plate production risk analysis model under each sub-category using the first feature set under the sub-category; A correlation model building module, used for training the correlation analysis model of the reinforcing plate production risk; The risk prediction module is used to input the real-time collected and pre-processed production data into the trained risk analysis models to obtain preliminary risk probabilities. The preliminary risk probabilities are adjusted based on the output of the association analysis model, and the adjusted quality defect probability or equipment failure probability of the current production process is output as the final risk probability. The early warning module is used to trigger the early warning module when the final risk probability exceeds the early warning threshold.

2. The reinforcing plate production risk perception and identification system according to claim 1 is characterized in that: The production equipment operation data includes equipment speed, temperature, vibration frequency, and operation time; the raw material quality data includes the strength, toughness, and chemical composition of the raw materials; the production process parameter data includes processing temperature, pressure, time, and processing speed; the operator behavior data includes the compliance of operation steps, operation time, and operation frequency; the production environment data includes ambient temperature, humidity, and dust concentration; the quality inspection results include defect type, location, and size; The preprocessing includes performing spatiotemporal alignment on the collected multi-source heterogeneous data, and constructing a production process time series data set in a unified format through missing value filling, outlier removal, noise filtering and standardization.

3. The reinforcing plate production risk perception and identification system according to claim 1 is characterized in that: The classification module implementation steps include: The first sub-category is divided into the types of defects; including cracks, deformation, substandard hardness and dimensional deviation; The second sub-category is divided according to the type of equipment anomaly; including mechanical failure, electrical failure, and sensor failure; Under each sub-category, the mean and standard deviation of multiple defective or abnormal samples of each parameter feature are calculated as the first mean and first standard deviation; The mean and standard deviation of each parameter characteristic when the quality inspection is passed and the equipment is operating normally are used as the second mean and second standard deviation; A first feature set of a sub-category is determined according to a first mean, a first standard deviation, a second mean, and a second standard deviation of a plurality of parameter features under the sub-category.

4. The reinforcing plate production risk perception and identification system according to claim 3 is characterized in that: The step of determining a first feature set of a sub-category according to a first mean, a first standard deviation, a second mean, and a second standard deviation of a plurality of parameter features under a sub-category comprises: Under each sub-category, the deviation ratio of the first mean value of each parameter relative to the second mean value is calculated as the first deviation; the deviation ratio of the first standard deviation of the parameter relative to the second standard deviation is calculated as the second deviation; Sort the multiple first offsets from high to low, and calculate the sum of the multiple first offsets; Calculating the sum of all sorted first offsets and multiplying the sum by a first preset proportional coefficient to obtain a first threshold; wherein the first preset proportional coefficient decreases as the number of defect samples increases; Select the first H parameter features from the sorted first offsets, so that the sum of the first offsets of the H parameter features reaches or exceeds the threshold for the first time, and use the first H parameter features as the first candidate features of the sub-classification; The first candidate feature and the second offset are combined to determine a final first feature set.

5. The reinforcing plate production risk perception and identification system according to claim 4 is characterized in that: The step of combining the first candidate feature and the second offset to determine a final first feature set includes: Calculating the sum of the second offsets of the parameters in the first candidate feature and multiplying the sum by the second preset proportional coefficient to obtain a second threshold; Set the first offset threshold coefficient α, and calculate α times the first threshold as the third threshold; where 0<α<1, Removing parameters from the first candidate feature that simultaneously satisfy the first offset being less than the third threshold and the second standard deviation being greater than the second threshold; The remaining parameter features constitute the final first feature set; When the candidate features do not include the key process parameters, the key process parameters are forcibly added to the first feature set.

6. The reinforcing plate production risk perception and identification system according to claim 4 is characterized in that: The system also performs the following steps: Perform correlation analysis on the first candidate feature set and remove redundant features; The bootstrap method was used to verify the stability of the feature screening results; The first proportional coefficient, the second proportional coefficient, and the threshold coefficient are updated periodically according to new data.

7. The reinforcing plate production risk perception and identification system according to claim 3 is characterized in that: The first mean is a weighted mean of values ​​of multiple defective or abnormal samples, where the weights are determined as follows: The time weight is calculated based on the test time of the defect or abnormal sample. The closer the test time is to the current time, the higher the time weight is. The grade weight is calculated based on the severity of the defect sample or anomaly. The higher the severity, the higher the grade weight. The grade weight is determined by the proportion of defect sizes exceeding a threshold; The final weight of the sample data is determined based on the time weight and level weight.

8. The reinforcing plate production risk perception and identification system according to claim 1 is characterized in that: The association model building module includes: Construct defect-fault co-occurrence labels based on historical data; calculate the conditional probabilities of different defect types and fault types; construct a defect-fault coupling matrix; the conditional probability includes: P(D i |F j ) and P(F j |D i ); where F j For the jth type of equipment failure, D i is the i-th type of quality defect; Use spatiotemporal graph convolutional networks to model device topology and process links, and output cross-modal correlation strength coefficient β ij ∈[0,1].

9. The reinforcing plate production risk perception and identification system according to claim 8, characterized in that: The risk prediction module implementation steps include: Input the real-time collected and pre-processed production data into the trained risk analysis models to obtain the preliminary risk probability P ini (D i ) and P ini (F j ); Adjust the preliminary risk probability based on the conditional probability and correlation strength coefficient output by the correlation analysis model; When a device failure is detected F j The initial risk probability is P ini (F j ), the corresponding quality defect D i The risk probability is adjusted to P adj (D i ): P adj (D i )=P ini (D i )+β ij ·P ini (F j )·P(D i |F j ) When quality defects are detected D i The initial risk probability is P ini (D i ), the corresponding equipment fails F j The risk probability is adjusted to P adj (F j ): P adj (F j )=P ini (F j )+β ij ·P ini (D i )·P(F j |D i ) Output the adjusted quality defect probability or equipment failure probability in the current production process as the final risk probability.

10. The reinforcing plate production risk perception and identification system according to claim 1, characterized in that: The steps for implementing the early warning module include: When the final risk probability exceeds the warning threshold, the warning module is triggered; the warning level is determined based on the risk probability and the correlation strength coefficient; Among them, the warning threshold is dynamically adjusted according to the historical frequency of defects.

Citation Information

Cited By

  • Casting product quality inspection monitoring method and system

    CN121073306A

  • A method and system for quality inspection and monitoring of casting products

    CN121073306B

  • LCD display screen production risk early warning method based on data analysis

    CN121279816A