A frame production quality data multidimensional analysis and abnormal fluctuation early warning system

Abstract

This invention discloses a multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data, specifically relating to the field of data analysis and early warning. The system includes: a data acquisition and standardization module that collects three types of raw data from chassis production: geometric dimensions, mechanical properties, and surface and weld images. After preprocessing, a standardized dataset and a benchmark parameter set are formed. A multi-dimensional analysis module sequentially performs quantitative assessments of dimensional consistency, structural stability, and surface integrity, outputting three sets of single-dimensional anomaly indices. A multi-mode coupling module integrates the single-dimensional indices into three sets of comprehensive coupled indices using three methods: linear weighting, nonlinear interaction, and extreme value sensitivity. A fusion decision module normalizes the coupled indices and then uses dynamic confidence weighting to obtain a single abnormal fluctuation determination value. An early warning execution module completes the three-level early warning threshold calibration iteration and triggers graded early warning handling based on the abnormal fluctuation value, achieving closed-loop control of chassis production quality anomalies throughout the entire process.

Description

Technical Field

[0001] This invention relates to the field of data analysis and early warning technology, and more specifically, to a multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data. Background Technology

[0002] As the automotive industry evolves towards electrification, intelligence, and lightweighting, the chassis, as the core load-bearing structure for vehicle safety and performance, is undergoing a systematic restructuring in its manufacturing logic and quality control model. Currently, chassis production quality control is gradually moving towards data-driven and intelligent approaches, with increasingly richer dimensions of data collection, now encompassing data related to geometric dimensions, mechanical properties, surface and weld images, and more during the production process.

[0003] Digital inspection equipment is gradually replacing traditional manual inspection methods, enabling precise capture and comprehensive inspection of key features of the chassis. The application of various sensors is becoming increasingly widespread, significantly increasing the amount of data generated per vehicle. Leading companies are actively promoting the construction of modular platforms, building a data closed loop of "platform-process-supply chain," and driving the systematic upgrading of quality control processes.

[0004] The chassis quality data monitoring market is expanding rapidly, related technologies are constantly iterating, and policy guidance and industrial practice are jointly promoting the gradual improvement of the industry standardization system, gradually forming a data-driven and intelligent quality control development trend covering the entire chassis production process, meeting the diversified quality requirements of vehicle durability, safety and adaptability to new energy vehicles.

[0005] However, it still has some drawbacks in practical use, such as:

[0006] 1. Existing frame production quality control technology lacks unified and standardized processing for various types of data such as geometric dimensions and mechanical properties. Data from different sources has messy formats and inconsistent numerical ranges, making it impossible to achieve efficient integration and collaborative analysis. This results in low data utilization, difficulty in uncovering the quality correlation patterns behind the data, and an inability to provide accurate and consistent data support for quality assessment, thus limiting the accuracy of control.

[0007] 2. Existing technologies for assessing frame quality often focus on a single dimension, failing to comprehensively cover key quality-influencing aspects such as size, structure, and surface. The assessment perspective is one-sided, unable to fully reflect the overall quality status of the frame, and is prone to overlooking potential quality hazards. It is difficult to meet the needs of high-end frames and new energy vehicles for refined and comprehensive quality control.

[0008] 3. Existing technologies have a single coupling analysis method for multi-dimensional quality data, lack diverse coupling logic, cannot adapt to the quality risk characteristics under different production conditions and process parameters, are difficult to accurately depict the superposition effect of multi-dimensional quality anomalies, and have a weak ability to identify hidden and complex quality risks.

[0009] 4. The warning thresholds of existing frame quality warning technologies are mostly set manually and cannot be dynamically adjusted according to changes in production processes, core equipment status, and raw material types. This makes them prone to false alarms and missed alarms. Furthermore, there is no clear graded warning and handling mechanism, and the response to different levels of quality anomalies is simplistic, resulting in low management efficiency. Summary of the Invention

[0010] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data, which solves the problems mentioned in the background art through the following solutions.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data, comprising:

[0012] Data acquisition and standardization module: Collects three types of raw data during the frame production process: geometric dimensions, mechanical properties, and surface and weld images. After time alignment, feature and benchmark parameter extraction and calibration, and normalization processing, a standardized dataset and benchmark parameter set are formed.

[0013] Multi-dimensional analysis module: Based on the standardized dataset and benchmark parameter set, quantitative evaluation is carried out dimension by dimension in the order of size consistency dimension, structural stability dimension and surface integrity dimension, and three sets of interval unified single-dimensional anomaly index are output.

[0014] Multi-coupling module: Based on the three sets of unified single-dimensional anomaly indices, three coupling methods are used in sequence: linear weighting, nonlinear interaction, and extreme value sensitivity, to integrate the single-dimensional anomaly into three sets of comprehensive coupling indices.

[0015] Fusion decision module: Normalizes the three sets of comprehensive coupling indicators, calculates the dynamic confidence of each coupling mode, and uses the dynamic confidence as a weighting coefficient to perform weighted fusion of the three sets of comprehensive coupling indicators, converging into a single abnormal fluctuation determination value;

[0016] Early warning execution module: calibrates the three-level early warning thresholds, and triggers graded early warning handling based on the single abnormal fluctuation determination value, establishes a dynamic iteration mechanism for the thresholds, and forms a closed loop for quality risk control.

[0017] The technical effects and advantages of this invention are as follows:

[0018] 1. This solution can standardize various quality data in the chassis production process. Through a unified time sequence alignment, normalization and parameter calibration process, it can achieve standardization and unification of data from different sources and dimensions, greatly improve data utilization, fully explore the quality correlation information between data, and provide high-quality and highly consistent data support for subsequent quality analysis and early warning, adapting to the needs of refined management and control.

[0019] 2. This solution adopts a multi-dimensional quality assessment logic, comprehensively covering key aspects such as frame size consistency, structural stability, and surface integrity. It conducts a comprehensive and accurate assessment of the overall quality status of the frame from multiple perspectives, effectively avoiding the omission of quality hazards caused by single-dimensional assessment, and improving the comprehensiveness and accuracy of quality assessment.

[0020] 3. This solution sets up a variety of flexible coupling analysis methods, which can be adapted to the quality risk characteristics under different production conditions. It can accurately depict the superposition effect of multi-dimensional quality anomalies and effectively identify hidden and complex quality risks. Compared with a single coupling method, it has stronger adaptability and more comprehensive risk identification.

[0021] 4. This solution adopts a fully data-driven dynamic threshold calibration and iteration mechanism, which can dynamically adjust the early warning threshold according to actual changes in production processes, equipment status, etc., to reduce false alarms and missed alarms. At the same time, a graded early warning and handling mechanism is set up to implement differentiated responses to quality anomalies of different degrees, so as to achieve accurate identification, efficient handling and closed-loop management of quality anomalies, and improve overall management efficiency. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0023] Figure 2 This is a schematic diagram of the data acquisition and standardization module of the present invention.

[0024] Figure 3 This is a schematic diagram of the multi-dimensional analysis module of the present invention.

[0025] Figure 4 This is a schematic diagram of the multi-coupling module of the present invention.

[0026] Figure 5 This is a schematic diagram of the fusion decision-making and early warning execution module of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] refer to Figure 1 - Figure 5 The system shown is a multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data, including:

[0029] Data acquisition and standardization module: Following the closed-loop steps of multi-source raw data acquisition, data preprocessing and standardization, the module first acquires three types of raw signals—geometric dimensions, mechanical properties, and surface and weld images—through acquisition equipment. Then, it performs time-axis correlation on heterogeneous data, extracts feature variables and reference benchmarks, and finally eliminates dimensional differences through normalization processing to form a standardized feature dataset and a fixed benchmark parameter set with unique identifiers.

[0030] It should be further explained that multi-source raw data collection specifically includes:

[0031] Raw geometric dimension data: acquired via an online 3D scanner (accuracy ≥ 0.01mm). The acquisition is triggered by a signal indicating that the frame has been machined / assembled and is in place. The acquisition frequency is synchronized with the production cycle. The acquisition scope includes: the 3D spatial coordinates of all key mounting holes on the frame, the contour dimensions of all key cross-sections of the frame, the key length / width / height parameters of the frame assembly, and the geometric shape and position parameters of key connection parts of the frame. The acquired data is stored in 3D point cloud format, with a unique identifier including the production cycle and frame number.

[0032] Raw mechanical performance data: Data is collected by force sensors (accuracy ≥ 0.1N) and strain gauges (accuracy ≥ 1με) installed on key parts of the welding robot spindle / clamp. The data acquisition is triggered by the start of the welding process and the acquisition frequency is 100Hz (real-time synchronous welding process). The acquisition range includes: real-time changes in the spindle pressure of the welding robot during the welding process, real-time detection data of the welding process and residual stress after welding at key connection points of the frame, and real-time data of clamping force of the clamp. The acquired data is stored in time-series curve format and includes a unique identifier such as welding process number, frame number, and acquisition timestamp.

[0033] Raw data for surface and weld seam images: Acquired using a high-resolution industrial camera (resolution ≥ 20 million pixels) and a linear scanning camera (scanning accuracy ≥ 0.05 mm). Acquisition is triggered by the completion of paint / coating treatment of the chassis and the completion of welding procedures. The acquisition frequency is synchronized with the production cycle. The acquisition range includes: high-definition images of the paint / coating on the entire chassis surface, linear scanning images of all major weld seam areas of the chassis, and high-definition images of defects in key surface parts of the chassis. The acquired images are in distortion-free bitmap format, with the shooting location, chassis number, and unique identifier of the production cycle. The image pixels are calibrated proportionally to the actual size.

[0034] It should be further explained that data preprocessing and standardization include: uniformly and standardized preprocessing of all collected multi-source heterogeneous raw data; extraction, calibration, and solidification of feature parameters and benchmark parameters required by subsequent analysis modules; unifying the data dimensions of the entire system and converting them into standardized feature data in the range [0, 1]; retaining unique identifiers such as chassis number, production cycle time, and timestamp throughout the preprocessing process to achieve full traceability of raw data, preprocessed data, and standardized data; and finally outputting standardized feature data and calibration benchmark parameters that meet the requirements of all subsequent analyses. The specific process is as follows:

[0035] Time alignment: Based on the production cycle number and unique timestamp, all raw data and extracted feature data of the three dimensions of geometric dimensions, mechanical properties, and surface and weld images are uniformly time-axis aligned to ensure that all data under the same frame and the same production cycle can be analyzed in a related manner. After alignment, all data are accompanied by a unified production cycle + frame number dual identifier.

[0036] Feature parameter extraction and baseline parameter calibration: From the raw data before standardization across the three dimensions, all feature parameters required for subsequent analysis are extracted, and all baseline reference parameters are calibrated; specifically as follows:

[0037] Geometric dimensions:

[0038] Extract the current size feature vector x: Extract the size deviation values ​​of key mounting holes, cross-sectional contours and assembly shapes from the original 3D point cloud data of geometric dimensions, and arrange them in a fixed dimension as the current size feature vector x of the frame. The dimension is consistent with the number of key dimensions collected.

[0039] Calibrate the mean vector μ of historical qualified samples: Collect raw data of qualified geometric dimensions of at least 1000 sets of frames produced under normal conditions, extract the dimension feature vectors after data cleaning, and calculate the mean of all vectors to obtain the mean vector μ;

[0040] The covariance matrix S of historical qualified samples is calculated using statistical analysis tools based on the above 1000 sets of grid size feature vectors, reflecting the correlation between the size features of qualified samples.

[0041] Mechanical properties:

[0042] Extracting real-time curves of welding pressure From the raw mechanical performance data, the real-time variation time-series data of the welding robot spindle pressure during the welding process of a single vehicle frame is extracted and arranged in time series to form the real-time curve of the current welding pressure. ;

[0043] Calibration of standard welding pressure profile Collect real-time spindle pressure curves from at least 500 sets of qualified welding processes on the vehicle frame. After data cleaning, obtain a standard template curve for welding pressure through curve fitting. This serves as a benchmark for consistency in the welding process;

[0044] Extracting residual stress values ​​from key measuring points From the raw mechanical performance data, extract the actual measured values ​​of residual stress after welding at n key connection points of the frame, denoted as . , ,..., (i=1, 2, ..., n, where n is the number of key measuring points, which is fixed to the process requirement value);

[0045] Calibrate the ideal value of residual stress at key measuring points Based on the chassis design standards and material mechanical property parameters, and through finite element simulation calculations combined with at least 300 sets of process verification data, the ideal reference values ​​for residual stress at each key connection point of the chassis were determined. ;

[0046] Surface and weld seam image categories:

[0047] Extracting original features of surface defects: Extracting original features of surface defects such as scratch area, scratch depth, stain area, and coating peeling area from high-definition images of surface paint film / coating. All features are converted into actual physical quantities according to the image pixel-actual size calibration relationship.

[0048] Extracting original features of weld defects: Extracting original features of weld defects such as number of pores, pore diameter, undercut length, undercut depth, and incomplete penetration length from the linear scan image of the weld area. All features are converted into actual physical quantities according to the image pixel-actual size calibration relationship.

[0049] It should be further explained that the normalization process transforms all extracted feature parameters into standardized feature data in the [0, 1] interval using the extreme value normalization method, eliminating the influence of dimensional differences on subsequent analysis. The normalization formula is unified as follows: Where: X represents the original value of the feature parameter to be normalized. and The upper and lower limits of the specifications are determined based on at least 1,000 sets of historical qualified data. After data cleaning, the historical maximum and minimum values ​​of the characteristic parameters are used for calibration, and the data is iteratively updated every 3 months based on new normal production data. All normalized standardized characteristic data are accompanied by the chassis number and production cycle number as dual identifiers.

[0050] Multi-dimensional analysis module: Based on the standardized features and fixed benchmarks output by the data acquisition and standardization module, it calculates dimensions one by one according to the dimensional consistency assessment, structural stability assessment, and surface integrity assessment. Each dimension follows a unified process of parameter substitution, formula calculation, full data-driven coefficient calibration, and result normalization mapping. It completes Mahalanobis distance calculation, DTW and stress deviation fusion, and defect score weighted summation, and finally outputs three sets of independent and interval-uniform dimensional anomalies. , , ;

[0051] It should be further explained that the evaluation of the dimensional consistency dimension is based on the geometrically standardized feature data and calibrated x, μ, and S output by the data acquisition and standardization module. The Mahalanobis distance algorithm, which considers the correlation between features, is used to quantify the deviation of the current batch of frame dimensional features from the distribution center of historical qualified samples. After normalization, the deviation is mapped to the interval [0, 1] to obtain the dimensional consistency anomaly value. , A higher value indicates a higher risk of dimensional inconsistencies; specifically as follows:

[0052] The degree of dimensional deviation is measured using the Mahalanobis distance, and its specific mathematical function is as follows:

[0053]

[0054] Substitute the current frame's dimensional feature vector x, the calibrated historical qualified sample mean vector μ, and the covariance matrix S; calculate the inverse of the covariance matrix S. ; Calculate the vector difference The product of the transpose and the inverse matrix Then the difference between the vectors Multiplying yields a scalar value, and taking the square root gives the original Mahalanobis distance. Mapping this original value to the [0, 1] interval using extreme value normalization yields the final size consistency anomaly D1. The normalized value... and Take the historical maximum and minimum values ​​of the Mahalanobis distance calculation results.

[0055] The benchmark parameter update rule is as follows: after each batch of production is completed, the new qualified frame size feature vector is included in the sample library, and the mean vector and covariance matrix are iteratively updated to ensure that the benchmark parameters match the current production process status.

[0056] It should be further clarified that the assessment of structural stability is based on the standardized mechanical performance characteristic data output by the data acquisition and standardization module, and the calibrated... , , , Based on this, the calculation results of both Dynamic Time Warping (DTW) distance and residual stress deviation are integrated, and weighting coefficients are used to determine the optimal approach. The evaluation weights of balancing welding process consistency and structural stress state are normalized to obtain the structural stability anomaly value. , A larger value indicates a higher risk of structural instability; its specific mathematical function is as follows:

[0057]

[0058] The specific calculation process is as follows:

[0059] DTW distance calculation: Real-time curve of current welding pressure Compared with standard welding pressure curve Perform dynamic time warping and calculate the similarity distance between the two time series curves. The larger the distance value, the worse the consistency of the welding process;

[0060] Calculation of average deviation of residual stress: Substitute the actual values ​​of residual stress at n key measuring points. Compared with ideal value Calculate the average absolute value of the deviation at each measuring point. The larger the value, the more the residual stress state deviates from the ideal value;

[0061] Coefficient weighted fusion: After normalizing the DTW distance value and the average deviation value of residual stress to the [0, 1] interval, the coefficients are weighted and then fused. With (1− Weighted fusion is performed to obtain the original fused value;

[0062] Final normalization: The original values ​​are mapped to the [0, 1] interval using the extreme value normalization method to obtain the final structural stability anomaly degree. .

[0063] in The proportion of welding process consistency in structural stability assessment, (1− The percentage represents the proportion of residual stress state in the structural stability assessment. ∈[0,1], the proportion perfectly matches the quantitative contribution of the two types of factors to structural quality failures; by collecting complete data on at least 500 sets of frame structural quality failures, two types of structural failures caused by inconsistent welding processes and abnormal residual stress were screened out, and other irrelevant failures were eliminated; the quantitative contribution of the two types of failures was calculated, and the contribution was... The total quality loss includes quantifiable economic costs such as rework costs, scrap costs, process adjustment costs, and after-sales quality losses, and is calculated based on the actual amount incurred; the quantifiable contribution of welding process inconsistencies is... The calibration value, the fault quantification contribution of residual stress anomaly is (1− The calibration value is determined by incorporating new structural failure data every 6 months, recalculating the quantitative contribution, and iteratively updating the data. The values ​​are set to ensure that the weights are highly consistent with the actual quality loss patterns caused by production failures.

[0064] It should be further explained that the assessment of the surface integrity dimension is based on the original physical characteristics of surface and weld defects output by the data acquisition and standardization module as input. A lightweight convolutional neural network combined with classical image processing is used to first complete the surface defect severity score based on the quantitative scoring criteria. Weld defect severity rating The calculation is then performed, and the two types of defect scores are integrated through the weighting coefficient β to directly obtain the surface integrity anomaly value in the [0,1] interval. , A higher value indicates a higher risk of surface integrity abnormalities;

[0065] The severity scoring of surface and weld defects is calculated using the original physical features of surface and weld defects extracted by the data acquisition and standardization module as input. A fully data-driven quantitative scoring system is constructed, and the scoring is performed using a lightweight MobileNetV2 convolutional neural network, outputting a score in the range [0, 1]. and The specifics are as follows:

[0066] Model selection and applicable scenarios: The MobileNetV2 lightweight convolutional neural network is selected, combined with traditional image processing algorithms such as edge detection, threshold segmentation, and contour extraction. The original physical features of the defects are used as the model input, and the quality loss quantification score is used as the model output. The inference time of a single frame image is ≤200ms, which meets the efficiency requirements of real-time detection and analysis on the chassis production line.

[0067] Surface defect rating ( Based on the vehicle frame surface paint / coating process standards, at least 500 sets of historical surface defect data were collected. The total mass loss corresponding to each defect's physical quantity was calculated. The mass loss of a single defect / the maximum mass loss of the surface category was used as the basic loss ratio for that defect. Then, a weighted sum was used to obtain the comprehensive loss ratio of surface defects, which is the percentage of total loss. The original scores are ultimately mapped to the [0, 1] interval; weighting formula: ,in, Assigning weights to the quality losses of each surface defect. =1, determined by the loss percentage based on historical fault data, coating peeling. =0.4, scratch depth =0.3, scratch area =0.2, stain area =0.1, For the actual quality loss of a single type of defect, This represents the maximum historical quality loss due to surface defects.

[0068] Weld defect scoring ( Based on the chassis welding process acceptance standards, at least 500 sets of historical weld defect data were collected. The total quality loss corresponding to each defect's physical quantity was calculated. The ratio of the single defect quality loss to the maximum quality loss of the weld category was used as the basic loss percentage for that defect. Then, a weighted summation was performed to obtain the comprehensive weld defect loss percentage, which is... The original scores are ultimately mapped to the [0, 1] interval; weighting formula: ,in, Assign a weight to the quality loss of each weld defect. =1, determined by the percentage of losses from historical fault data, indicating incomplete penetration. =0.4, bite depth =0.3, pore diameter =0.2, number of pores =0.1, For the actual quality loss of a single type of defect, This represents the maximum historical quality loss due to weld defects.

[0069] Model training and real-time scoring calculation process:

[0070] Sample labeling: Substitute the extracted original physical quantities of surface / weld defects into the above scoring criteria to calculate the true quality loss score (0-1) for each sample, which serves as the label for model training;

[0071] Model Training: The quantized values ​​of the original defect features are used as the model input, and the true quality loss score is used as the output label. The training and validation sets are divided in an 8:2 ratio. The MobileNetV2 model is trained using gradient descent, and handcrafted features from traditional image processing are incorporated to improve the model's robustness. Training continues until the goodness of fit on the validation set is achieved. ≥0.9 after curing model;

[0072] Real-time scoring: Input the original physical features of the surface / weld defects of the frame to be inspected into the solidified model. The model automatically completes feature extraction and quantification calculation, and directly outputs the results. , The closer the score is to 1, the higher the severity of the defect and the greater the risk of quality loss.

[0073] Scoring Iteration: If the deviation rate between the model output score and the actual quality loss is >3%, new defect samples and quality loss data are immediately added, the model is retrained and the scoring standard weights are updated, and the process is carried out on a regular basis once every quarter.

[0074] Surface integrity anomaly The model output , Substituting directly into the function for weighted summation, since the two scores have been normalized to the interval [0, 1], the summation result is directly obtained. ;in, The weight of the influence of surface defects on surface integrity, (1- ) represents the influence weight of weld defects. The weights are objectively determined by the quality loss and failure frequency of the two types of defects, ∈[0,1]. A dual-dimensional quantitative statistical method based on quality loss and failure frequency is used, with the following steps:

[0075] Collect no fewer than 500 sets of complete data on surface and weld quality defects, and statistically analyze the failure frequency and corresponding total quality loss of the two types of defects.

[0076] Calculate the comprehensive quantification coefficient: The coefficient weights are determined by data correlation analysis, and the correlation coefficient with actual quality risk is ≥0.85;

[0077] The comprehensive quantification coefficient of surface defects is The calibration value, the comprehensive quantitative coefficient of weld defects is (1− Calibration values; new fault data is used to re-calculate and iteratively update the data every quarter. Values.

[0078] Multi-coupling module: Analyzing the output of the multi-dimensional analysis module , , As input, the calculations are performed sequentially in the order of linear weighted coupling, nonlinear interactive coupling, and extreme value sensitive coupling. For each coupling method, the model coefficients, weights, and adjustment parameters are first calibrated using full data, followed by formula calculations and interval mapping. Multi-dimensional anomaly information is integrated from three independent perspectives: average anomaly level, risk amplification effect, and bottleneck risk. Finally, three sets of comparable comprehensive coupling indicators within the same interval are output. , , .

[0079] It needs to be further explained that, To achieve linear weighted coupling and reflect the average anomaly level, based on the quantitative logic that anomalies in each dimension independently contribute to the overall quality risk, a fully data-driven combined weighting method is used to determine the dimension weights, and a comprehensive coupling index is obtained through weighted summation. This is used to characterize the average level of anomalies in the overall quality of the chassis. The higher the value, the higher the overall average risk of anomalies across all dimensions; the specific mathematical function is as follows:

[0080] ; ;

[0081] Substitute the dimensionality anomaly from the multidimensional analysis module output. , , and the corresponding weights of the fully data-driven calibration. , , Calculate the product of the anomaly score and its corresponding weight for each dimension, and sum the products to obtain the result. Original calculated values; where weights The fully data-driven value retrieval method is as follows:

[0082] , , The weights for quantifying the contribution of dimensional consistency, structural stability, and surface integrity anomalies to the overall quality risk of the chassis are defined, with the weight values ​​positively correlated with the quality loss and failure frequency caused by that dimension. A combination of quality impact quantification and entropy weighting is used for weighting.

[0083] Collect at least 500 sets of complete frame quality failure and normal production data, and calculate the total quality loss and failure frequency corresponding to anomalies in each dimension;

[0084] A two-dimensional quantitative statistical method, namely quality loss and failure frequency, is used to calculate the weight of the quality impact of each dimension. ;

[0085] Based on normal production samples , , The sequence is used to calculate the objective weights of pure data using the entropy weight method. ;

[0086] The two weights are combined in a 1:1 ratio: =0.5 +0.5 And perform normalization processing to ensure Every 6 months, new production and failure data are included, and the weights are recalculated and iteratively updated.

[0087] It needs to be further explained that, To address the nonlinear interactive coupling and reflect the risk amplification effect, based on the engineering principle that the superposition of multi-dimensional anomalies leads to nonlinear risk amplification, a multinomial model containing single-dimensional terms and two-dimensional interaction terms is constructed. All coefficients are calibrated through nonlinear regression of historical fault data, and a comprehensive coupling index is calculated. It is used to characterize the level of risk amplification during multi-dimensional abnormal concurrency. The larger the value, the more significant the risk amplification effect caused by the interaction between dimensions; the specific mathematical function is as follows:

[0088] ;

[0089] Substitution , , Single-dimensional coefficients , , With interaction coefficient , , ; Calculate the sum of linear terms in a single dimension Characterize the basic risk contribution of single-dimensional anomalies; calculate the sum of two-dimensional interaction terms. This characterizes the nonlinear amplification contribution of simultaneous occurrence of two-dimensional anomalies; the sum of the two terms yields... The original values ​​are mapped to the interval [0, 1] using the extreme value normalization method to obtain the final value. ;in, This represents the basic contribution coefficient of a single-dimensional anomaly to the overall risk. This is a risk amplification coefficient representing the superposition of anomalies in two dimensions. The magnitude of the coefficient is determined by the quantitative correlation of historical failure data; it is determined based on multivariate nonlinear regression of historical failure data.

[0090] Collect at least 300 sets of valid fault data, including fault states. , , , and the overall quantitative risk value in the [0,1] interval obtained by mapping the total quality loss;

[0091] The dataset was divided into training and validation sets in an 8:2 ratio. , , , , , Using the overall quantitative risk value as the dependent variable, a regression model is constructed.

[0092] The gradient descent method was used to train the model to verify the set fit. ≥0.85 is the convergence criterion; the output model parameters after convergence are... , Each quarter, newly added fault samples are included for retraining, and all coefficients are updated iteratively.

[0093] It needs to be further explained that, To achieve extreme value-sensitive coupling and reflect the risk of bottlenecks, a computational model is constructed using the maximum value of dimensional anomaly as the core indicator, superimposed with an anomaly dispersion penalty term. Adjustment coefficients are determined through data simulation and experimental fitting, and a comprehensive coupling index is calculated. This is used to highlight the weakest dimension while also taking into account the overall volatility dispersion. The larger the value, the more prominent the quality shortcomings and the higher the abnormal dispersion between dimensions. The specific mathematical function is as follows:

[0094]

[0095] Substitution , , With adjustment coefficient Calculate the maximum value of the anomaly score across the three dimensions. Identify the weakest dimension in terms of current quality; calculate the standard deviation of the outlier rates of the three dimensions. Quantify the degree of dispersion of anomalies between dimensions; calculate the dispersion penalty term. Summing with the maximum value yields The original values ​​are mapped to the interval [0, 1] using the extreme value normalization method to obtain the final values. ;in This is the quantitative penalty coefficient for the risk of shortcomings due to dimensional anomalies in dispersion. The larger the value, the stronger the weighted impact of the dispersion of the abnormal distribution on the overall risk;

[0096] Value determination process: Determined through two-sample simulation fitting using process test data and historical production data.

[0097] Conduct no fewer than 20 sets of process simulation tests to obtain actual quality risk data under different abnormal dispersion conditions, and determine Initial interval [0.1, 0.3]; extract at least 500 sets of historical production data, iterate through different γ values ​​within the initial interval, and calculate the corresponding... Sequence; calculate each γ level The Pearson correlation coefficient with the actual fault label is used as the value corresponding to the maximum value of the correlation coefficient, and γ is refitted and fine-tuned every 3 months by combining the new process test and production data.

[0098] Fusion Decision Module: Three sets of comprehensive coupling indicators output by the multi-mode coupling module. , , Using the input as input, the calculation is completed in a progressive process of unified normalization of coupling indicators, dynamic confidence calculation and confidence weighted fusion. The comprehensive coupling indicators are converged into the abnormal fluctuation determination value F by using data statistical features to complete the confidence assignment and weighted calculation.

[0099] It should be further explained that the confidence-based adaptive fusion first performs unified interval normalization on the three sets of coupling indicators to eliminate the differences in numerical range caused by different coupling calculation logics. Then, based on the variance within the sliding time window, the dynamic confidence of each coupling method is calculated. The smaller the variance, the more stable the recent output of the coupling method is, and the higher the corresponding confidence. The specifics are as follows:

[0100] Coupling index normalization processing

[0101] right , , Using the same extreme value normalization method as the front-end module, it is uniformly mapped to the [0, 1] interval to obtain the normalized coupling index. , , Its normalized data function is as follows:

[0102] ;

[0103] in, This is the original comprehensive coupling index output by the multi-mode coupling module;

[0104] , Based on no fewer than 200 sets of continuous normal production data The cleaning results are calibrated using the historical minimum and maximum values, respectively.

[0105] Update rule: New production data will be included and statistics will be recalculated every month. and And complete the normalized benchmark iteration.

[0106] Dynamic confidence Calculation: Using the stability of the recent output of the coupling index as the sole criterion, a dynamic confidence calculation model is constructed, the specific mathematical function of which is:

[0107] The calculation process and rules are as follows:

[0108] Sliding time window setting: Based on the frame production cycle time, select the most recent 50 consecutive production cycles. As a sample for variance calculation, the window length is adjusted proportionally and synchronously when production cycle time is adjusted.

[0109] Recent variance calculation: within the calculation window Sample variance The smaller the variance, the better the output consistency and the higher the reliability of the coupling method.

[0110] Confidence level calculation: Substitute the variance into the formula to obtain the dynamic confidence level. The value range is (0, 1], and it is positively correlated with the output stability of the coupling method; the confidence level is updated in real time with the production data throughout the process, and no manual fixed value or correction item is set.

[0111] It should be further explained that the calculation of the abnormal fluctuation determination value uses dynamic confidence level as the weighting coefficient, and performs a weighted average of the normalized coupled indicators to fuse the three sets of multi-perspective coupled indicators into a unique abnormal fluctuation determination value F, where F∈[0,1]. The magnitude of F is positively correlated with the risk of abnormal quality in chassis production; the details are as follows:

[0112] The final formula for calculating the abnormal fluctuation determination value is: ;

[0113] Substitute the normalized coupling index , , With corresponding dynamic confidence level , , ;

[0114] The numerator is calculated as the sum of the products of each normalized coupling index and its confidence level. ;

[0115] Calculate the denominator: the sum of all dynamic confidence levels. , as a weighted normalization factor;

[0116] Dividing the numerator and denominator yields the final anomalous fluctuation determination value F. ∈[0,1]、 Since F ∈ (0, 1), it falls within the interval [0, 1] and does not require normalization.

[0117] The early warning execution module uses the unique abnormal fluctuation determination value F output by the fusion decision module as the core judgment basis. It follows a closed-loop logic of numerical hierarchical comparison, early warning level triggering, corresponding measure execution, and dynamic threshold iteration. It completes the full data-driven calibration of the three-level early warning thresholds through historical data statistical distribution and process capability index. It implements hierarchical early warning response based on the relationship between F and the threshold, and establishes a periodic automatic optimization mechanism for the threshold. This enables accurate identification, hierarchical handling, and dynamic adaptation of quality anomalies, forming a complete quality risk control closed loop of judgment-early warning-handling-optimization. This ensures stable production process quality and prevents the continuous generation and circulation of non-conforming products.

[0118] It should be further explained that the triggering mechanism for the Level 3 warning is based on the abnormal fluctuation determination value F and the Level 3 warning threshold. , , The numerical comparison results serve as the sole triggering condition, classifying risks into three tiers: low, medium-high, and extremely high. Correspondingly, standardized response measures are implemented with progressively stronger measures to achieve differentiated handling of anomalies at different risk levels; details are as follows:

[0119] Level 1 Warning (Attention Level): Triggering condition: F> ;

[0120] Response measures: The system automatically displays a yellow warning sign on the production quality monitoring dashboard, simultaneously indicating the corresponding chassis number, production cycle time, and real-time F value; it automatically pushes warning information to the production line shift leader, requiring them to conduct a preliminary inspection of the appearance and operating status of the corresponding production station and process parameters; the shift leader must complete the on-site inspection within 10 minutes and enter the results into the system. If there is no substantial abnormality, the system will automatically lift the warning; if there is an abnormal trend, it will continue to monitor and record data changes.

[0121] Level 2 Warning (Intervention Level): Triggering Condition: F> ;

[0122] Response measures: The system displays a red alarm icon on the monitoring dashboard and simultaneously activates the workshop's audible and visual warning devices; it automatically sends warning commands to process engineers and quality engineers, requiring them to immediately arrive at the corresponding production workstations to verify key aspects such as equipment operating parameters, welding processes, and the calibration status of testing equipment; it performs 100% full-volume strict inspection on the current production batch of products, distinguishing between qualified parts and parts awaiting re-inspection; engineers must complete a preliminary cause analysis within 30 minutes and upload the analysis results and handling suggestions to the system.

[0123] Level 3 Warning (Stop Level): Triggering Condition: F> ;

[0124] Response measures: The system automatically sends a shutdown command (or suggests shutdown) to the production control system to suspend the corresponding production process and prevent the continued production of defective products; simultaneously, it sends an emergency warning notification to the heads of the production, quality, and technical departments to initiate a multi-department joint investigation process; all frames of the corresponding production batch before the shutdown are subject to full batch isolation, labeling, and special inspection, and are strictly prohibited from entering the next process without passing the re-inspection; the joint investigation team must issue a preliminary abnormality investigation direction within 1 hour after the shutdown, and can apply for resumption of production only after the root cause has been located and rectified.

[0125] It should be further explained that the three-level early warning threshold is subject to dynamic management, and a three-level early warning threshold system is established. , , The fully data-driven calibration, verification, and iteration mechanism uses the statistical distribution of normal production data, false alarm rate, and false negative rate as the core quantitative basis to achieve dynamic matching of thresholds with production process status and equipment performance, avoiding false alarm and false negative problems caused by fixed thresholds; the details are as follows:

[0126] Initial calibration: The percentile method is used to complete the initial threshold calibration. Specific steps are as follows:

[0127] Collect at least 500 sets of abnormal fluctuation values ​​F under normal and stable production conditions of the chassis. Through data cleaning, remove non-process abnormal data such as equipment failure and human error to construct a pure statistical distribution of normal production F values.

[0128] Based solely on the percentile of the statistical distribution, an initial threshold is set: Take the 85th percentile. Take the 95th percentile. Use the 99th percentile to ensure that the initial threshold closely matches the distribution of normal production data.

[0129] Validation and Optimization: After the initial threshold is implemented, quantitative validation and automatic optimization will be carried out on a one-month cycle. The optimization rules are based entirely on statistical indicators.

[0130] The formulas for calculating the false alarm rate and the false negative rate within the statistical period are as follows:

[0131] False alarm rate = Number of warnings without actual quality issues / Total number of warnings;

[0132] False negative rate = Number of times an actual quality problem occurred but no warning was triggered / Total number of quality problems;

[0133] Optimization rule: If the false positive rate is greater than 5%, adjust the rate upwards by 5% accordingly. , , If the underreporting rate is greater than 1%, the rate will be adjusted downwards by 5%. , , ;

[0134] Repeatedly perform verification and optimization until the false alarm rate is ≤5% and the false negative rate is ≤1%, then solidify this threshold as the official operating threshold.

[0135] Dynamic updates: The system has online learning and automatic update capabilities, performing routine threshold iterations every 3 months. It also supports rapid recalibration after significant changes in production conditions, and the entire process is based solely on the statistical distribution of the F-value and adjustments to the early warning effect.

[0136] Extract the newly added normal production F-value data from the past three months, incorporate it into the historical sample library, re-clean and reconstruct the statistical distribution, and recalibrate the benchmark threshold according to the 85th, 95th and 99th percentiles to ensure that the threshold conforms to the long-term production data distribution pattern.

[0137] Simultaneously review the overall early warning accuracy within the period. If the average false alarm rate and false alarm rate deviate from the optimization target range, make a small adjustment within ±3% based on the percentile calibration results to maintain the accuracy of the early warning.

[0138] When there are significant changes in production processes, core testing equipment, or raw material types, immediately re-collect 500 sets of F-value data for normal production after the changes, re-execute the entire process of initial calibration, verification, and optimization, and generate a new threshold adapted to the new production state.

[0139] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0140] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A frame production quality data multi-dimensional analysis and abnormal fluctuation early warning system, characterized in that, include: Data acquisition and standardization module: Collects three types of raw data during the frame production process: geometric dimensions, mechanical properties, and surface and weld images. After time alignment, feature and benchmark parameter extraction and calibration, and normalization processing, a standardized dataset and benchmark parameter set are formed. Multi-dimensional analysis module: Based on the standardized dataset and benchmark parameter set, quantitative evaluation is carried out dimension by dimension in the order of size consistency dimension, structural stability dimension and surface integrity dimension, and three sets of interval unified single-dimensional anomaly index are output. Multi-coupling module: Based on the three sets of unified single-dimensional anomaly indices, three coupling methods are used in sequence: linear weighting, nonlinear interaction, and extreme value sensitivity, to integrate the single-dimensional anomaly into three sets of comprehensive coupling indices. Fusion decision module: Normalizes the three sets of comprehensive coupling indicators, calculates the dynamic confidence of each coupling mode, and uses the dynamic confidence as a weighting coefficient to perform weighted fusion of the three sets of comprehensive coupling indicators, converging into a single abnormal fluctuation determination value; Early warning execution module: calibrates the three-level early warning thresholds, and triggers graded early warning handling based on the single abnormal fluctuation determination value, establishes a dynamic iteration mechanism for the thresholds, and forms a closed loop for quality risk control.

2. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The feature and reference parameter extraction and calibration include: geometric dimension parameter extraction and calibration, mechanical property parameter extraction and calibration, and surface and weld image parameter extraction and calibration. The geometric dimension category extracts dimensional deviation characteristic parameters from the original geometric dimension data of the frame, and calibrates corresponding statistical reference benchmark parameters based on historical qualified dimensional data; the mechanical performance category extracts welding pressure timing characteristic parameters and residual stress characteristic parameters at each connection point from the original mechanical performance data of the frame, fits welding pressure standard template benchmark parameters based on qualified welding process data, and calibrates ideal reference benchmark parameters for residual stress based on frame design standards and process verification data; the surface and weld image category extracts surface defect characteristic parameters and weld defect characteristic parameters from the original surface and weld image data of the frame.

3. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The dimension of size consistency includes: The Mahalanobis distance algorithm is used to calculate the deviation of the frame size characteristics from the distribution center of historical qualified samples. The calculation results of the deviation are normalized and mapped to a unified numerical range to obtain the size consistency anomaly index. The statistical benchmark parameters of historical qualified samples used in the evaluation process are updated iteratively by incorporating newly added qualified frame size characteristic parameters into the sample library after each batch of production is completed.

4. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The structural stability dimension includes: Based on a standardized dataset and benchmark parameter set, welding pressure time-series characteristic parameters, residual stress characteristic parameters at each connection point, and calibrated welding pressure standard template benchmark parameters and residual stress ideal reference benchmark parameters are extracted. The dynamic time warping (DTW) distance between the current welding pressure time-series curve and the welding pressure standard template curve, as well as the average deviation between the actual residual stress value and the ideal reference value at each connection point, are calculated. After normalizing the DTW distance value and the average deviation value of residual stress, they are weighted and fused using weight coefficients. The fused result is then normalized and mapped to a unified numerical range to obtain the structural stability anomaly index. The weight coefficient represents the proportion of welding process consistency in the structural stability assessment. New structural fault data is incorporated every 6 months for recalculation, completing the iterative update of the weight coefficient value.

5. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The surface integrity dimension includes: Based on standardized datasets and benchmark parameter sets, surface defect feature parameters and weld defect feature parameters are extracted. A lightweight MobileNetV2 convolutional neural network combined with image processing algorithms is used to calculate the severity scores of surface defects and weld defects, respectively, with both scores mapped to a unified numerical range. The two scores are then weighted and summed using weight coefficients to obtain a surface integrity anomaly index. The lightweight convolutional neural network is trained with defect feature parameters as input and quality loss quantification scores as labels. After achieving a preset goodness of fit, the model parameters are fixed. New defect samples and quality loss data are included in the retraining every quarter. When the deviation rate between the model score and the actual quality loss exceeds a preset range, the model and weight parameters are updated.

6. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The linear weighting includes: Using three sets of interval-unified single-dimensional anomaly indices output by the multi-dimensional analysis module as input, a weighting method combining the quality impact quantification method and the entropy weight method is adopted to determine the weights corresponding to the three sets of single-dimensional anomaly indices, with the sum of all weights being 1. Each set of single-dimensional anomaly indices is multiplied by its corresponding weight, and all product results are summed to obtain a linearly weighted comprehensive coupling index.

7. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The nonlinear interaction includes: Using three sets of interval-unified single-dimensional anomaly indices output by the multi-dimensional analysis module as input, a multinomial coupled model containing single-dimensional linear terms and two-dimensional interaction terms is constructed. A full-data-driven approach using multivariate nonlinear regression of historical fault data is employed to calibrate the single-dimensional basic contribution coefficient and the two-dimensional interaction risk amplification coefficient in the model. During calibration, the three sets of single-dimensional anomaly indices and their products are used as independent variables, and the overall quantitative risk value corresponding to historical faults is used as the dependent variable. The model is trained by dividing the training and validation sets until a preset goodness of fit is achieved, after which the model parameters are solidified. The three sets of single-dimensional anomaly indices are substituted into the solidified multinomial model, and the sum of the single-dimensional linear terms and the sum of the two-dimensional interaction terms are calculated separately. The sums of the two terms are added to obtain the original coupled value. Through normalization, the original coupled value is mapped to a unified numerical interval to obtain the nonlinear interaction comprehensive coupled index.

8. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The extreme value sensitivity includes: Using three sets of interval-unified single-dimensional anomaly indices output by the multi-dimensional analysis module as input, the maximum value among the three sets of single-dimensional anomaly indices is first calculated, and then the standard deviation of the three sets of single-dimensional anomaly indices is calculated. The standard deviations are weighted by adjustment coefficients to obtain an anomaly dispersion penalty term. The maximum value of the three sets of single-dimensional anomaly indices is added to the anomaly dispersion penalty term to obtain the extreme value sensitive coupling original value. The coupling original value is mapped to a unified numerical interval through normalization processing to obtain the extreme value sensitive comprehensive coupling index.

9. The multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The weighted fusion of the three sets of comprehensive coupling indicators includes: The three sets of comprehensive coupling indices output by the multi-mode coupling module are normalized to obtain three sets of normalized coupling indices. Then, the dynamic confidence level corresponding to each set of normalized coupling indices is calculated. The dynamic confidence level is based on the variance of the comprehensive coupling indices within the sliding time window. The sliding time window is based on the frame production cycle time. The comprehensive coupling indices of the most recent 50 consecutive production cycles are selected as the variance calculation sample. The window length is adjusted synchronously when the production cycle time is adjusted. The dynamic confidence level is obtained by variance conversion and the value range is between 0 and 1. The dynamic confidence level corresponding to each set of normalized coupling indices is used as the weighting coefficient. Each set of normalized coupling indices is multiplied by its corresponding dynamic confidence level. The sum of all products is used as the numerator, and the sum of all dynamic confidence levels is used as the denominator. The final abnormal fluctuation determination value is obtained by dividing the numerator by the denominator.

10. A multi-dimensional analysis and abnormal fluctuation early warning system for chassis production quality data according to claim 1, characterized in that: The early warning execution module includes: threshold calibration and iteration, and hierarchical early warning triggering and response; Threshold calibration and iteration: Abnormal fluctuation values ​​are collected under normal and stable production conditions of the chassis. After data cleaning, the percentile method is used to set three-level initial warning thresholds. The false alarm rate and false alarm rate of the warning are statistically analyzed at a preset period and the thresholds are adjusted accordingly. Threshold iteration is performed at a fixed period. Tiered early warning triggering and response: The abnormal fluctuation value is compared with the three-level early warning threshold. Based on the comparison result, the corresponding level of early warning response is triggered, and the corresponding level of early warning sign display, information push, workstation verification, stricter inspection, process shutdown, batch isolation and joint investigation operations are performed. The early warning triggering information, handling process and handling results are recorded.

Images