A quality inspection method for automotive aluminum trim parts based on big data

Abstract

This invention discloses a big data-based quality inspection method for automotive aluminum trim parts, specifically relating to the fields of quality management and quality inspection decision-making. It addresses the problem that existing aluminum trim part quality inspection methods rely on fixed rules and experience-based judgments, making it difficult to dynamically optimize quality inspection performance and resource allocation by incorporating after-sales quality information. This method structures and classifies vehicle after-sales damage data, extracts key process features from production process time-series data, and constructs an early defect propagation model to predict the probability of different defect types. Based on this, it dynamically generates multi-level hierarchical quality inspection strategies according to defect risk levels, and integrates quality inspection execution costs and historical quality loss costs. Through multi-objective optimization, it outputs quality inspection performance indicators for the next production cycle, thereby improving the scientific nature and decision-making value of quality risk management while controlling costs, thus adapting to the management needs of both refined operations and cost constraints.

Description

Technical Field

[0001] This invention relates to the field of quality management and quality inspection decision-making technology, and more specifically, to a quality inspection method for automotive aluminum trim products based on big data. Background Technology

[0002] Current quality management of automotive aluminum trim parts mainly revolves around production execution and result inspection. Its core objective is to control the outflow of non-conforming products and reduce after-sales quality risks. However, in practice, there are still common problems such as management decisions relying on experience, rigid resource allocation, and insufficient utilization of cross-process information.

[0003] With the widespread adoption of large-scale and multi-batch parallel production models in automobile manufacturing, aluminum trim components exhibit varying quality performance under different production cycles and process conditions. Relying solely on fixed quality inspection rules or static sampling schemes makes it difficult to balance quality risk and quality inspection investment within cost constraints. Meanwhile, companies accumulate substantial damage and loss data during after-sales repairs, quality claims, and internal quality assessments. However, this data is primarily used for post-event statistics and responsibility analysis, lacking a systematic data processing mechanism for production and quality inspection decision-making. This results in a lack of foresight and overall coordination in quality management activities.

[0004] In quality management practice, how to implement differentiated testing strategies for products with different risk levels based on quality losses and production status, rationally allocate testing resources, and reduce potential quality risks while controlling execution costs has become a key issue in enterprise operation and management. Therefore, it is necessary to construct a data processing method for quality management and quality inspection decision-making. Through comprehensive analysis of quality events, production processes, and cost information, this method can achieve dynamic generation and performance optimization of quality inspection strategies, providing enterprises with more economical and decision-making-valuable quality management tools to meet the needs of quality inspection control that emphasizes both refined operations and cost constraints. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a big data-based quality inspection method for automotive aluminum trim products to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A quality inspection method for automotive aluminum trim parts based on big data includes the following steps:

[0008] S1. Extract damage records containing batch identifiers of aluminum trim parts from vehicle after-sales maintenance data sources, perform data cleaning on the damage records, and generate structured batch damage data.

[0009] S2. Perform unsupervised learning based on spectral clustering on batch damage data to classify the damage pattern groups of aluminum decorative parts.

[0010] S3. Based on the batch identifier of the aluminum trim parts, trace the time-series production process data of the corresponding production cycle, and cut it into production segment feature sequences using the sliding window method;

[0011] S4. For each damage pattern family, input the feature sequences of all batches of production fragments contained therein into the integrated classification framework to identify key process feature combinations.

[0012] S5. Based on the combination of key production process characteristics, construct an early propagation network for aluminum decorative parts process defects to predict the types of measurable defects and their probability of occurrence.

[0013] S6. Dynamically generate multi-level hierarchical quality inspection strategies based on the predicted defect types and occurrence probabilities.

[0014] S7. Integrate the estimated execution cost of multi-level hierarchical quality inspection strategies with historical quality loss cost data corresponding to each damage mode group, and output a set of quality inspection performance indicators for the next production cycle through a dynamic bi-objective optimization algorithm.

[0015] As a further aspect of the present invention, in step S1, generating structured batch damage data specifically includes:

[0016] Access vehicle after-sales maintenance data sources and extract damage records for aluminum trim parts. The records include batch identifiers of the aluminum trim parts and descriptions of the damage morphology.

[0017] Data cleaning is performed on the extracted damage records to unify the format of part batch identifiers and map the damage morphology descriptions to standardized damage morphology codes.

[0018] The damage records after cleaning are integrated according to the batch identifier of the aluminum trim parts to generate a structured damage dataset indexed by the part batch.

[0019] As a further aspect of the present invention, in step S2, classifying the damage mode families of aluminum trim parts specifically includes:

[0020] Based on the frequency of occurrence of various damage morphology codes of each batch of aluminum trim parts in the structured damage dataset, calculate the cosine similarity between any two batches of aluminum trim parts.

[0021] Construct a symmetric batch similarity matrix for parts and calculate the corresponding Laplacian matrix. Perform eigenvalue decomposition on the Laplacian matrix and select the eigenvectors corresponding to the first few largest eigenvalues ​​to form a new feature space.

[0022] In the new feature space, a clustering algorithm is used to divide all aluminum trim parts into several mutually exclusive sets. Each set is defined as a damage pattern family and labeled with a family identifier.

[0023] As a further aspect of the present invention, in step S3, tracing the time-series production process data corresponding to the production cycle based on the batch identifier of the aluminum trim parts, and cutting it into a production segment feature sequence using a sliding window method specifically includes:

[0024] Based on the batch identifier of aluminum trim parts, retrieve multi-dimensional time series data of each batch recorded along the complete production chain from historical production lines;

[0025] The retrieved multidimensional time series data are aligned and stitched together according to the manufacturing process nodes to form a continuous global production process data stream for this batch;

[0026] Based on the theoretical production cycle time of a single product, the width and step interval of the sliding window are defined, and the global production process data stream is continuously captured along the time axis.

[0027] For each data segment captured by the sliding window, calculate the statistical characteristics of various production parameters contained therein within the time window. Arrange the statistical characteristics of all windows in the batch in chronological order to form the production segment feature sequence corresponding to this batch.

[0028] As a further aspect of the present invention, in step S4, for each damage pattern family, the feature sequences of all batches of production segments contained therein are input into the integrated classification framework to identify key process feature combinations, specifically including:

[0029] For each damage mode family, the production segment feature sequence corresponding to the batch identifier of all aluminum trim parts in the family is used as the positive sample set. Normal parts batches that are not classified into any damage mode family are extracted from historical production data, and the corresponding production segment feature sequence is used as the negative sample set.

[0030] An ensemble classification model is trained using positive and negative sample sets, taking production fragment feature sequences as input and batch classification prediction as output.

[0031] The cumulative contribution of each process feature to the sample distinction result during the classification decision node division in the model training process is statistically analyzed. The importance ranking of the process features on which the integrated classification model depends in the classification decision is calculated, and the subset of features with the highest importance is selected as the key process feature combination.

[0032] As a further aspect of the present invention, in step S5, constructing an early propagation network for aluminum trim manufacturing defects and predicting the types and probabilities of measurable defects specifically includes:

[0033] Key process features are combined and mapped to corresponding manufacturing nodes. A directed graph network model is constructed based on the sequential relationship of manufacturing nodes. Nodes represent manufacturing operations that carry key process features, and edges represent the propagation relationship of defects along the process path.

[0034] By utilizing the co-occurrence relationship between the same key process features extracted from historical production data and the final detected defects, the conditional probability of defect state transition between adjacent nodes in the directed graph network model is calculated.

[0035] In the current batch of aluminum decorative parts, the feature sequence of production segments containing key process characteristics is used as input. Forward propagation reasoning is performed along the process path in the directed graph network model to calculate the probability of various defects that are finally detected in each downstream process in real time. The defect types with probabilities exceeding the set confidence level and their corresponding probability values ​​are used as prediction outputs.

[0036] As a further aspect of the present invention, in step S6, dynamically generating a multi-level hierarchical quality inspection strategy based on the predicted defect type and probability of occurrence specifically includes:

[0037] The probability of occurrence for each predicted defect type is divided into different risk level ranges;

[0038] For each risk level range, a preset detection intensity level is matched. The detection intensity level defines the detection type and sampling benchmark ratio corresponding to the risk level defect, and at the same time determines the quality inspection process to be performed for the corresponding detection intensity level.

[0039] By integrating the risk level, inspection intensity level, and quality inspection information corresponding to all predicted defect types in the current batch of aluminum trim parts, a dynamic quality inspection checklist with multi-level hierarchical quality inspection strategies is generated.

[0040] As a further aspect of the present invention, in step S7, outputting the set of quality inspection performance indicators for the next production cycle specifically includes:

[0041] Based on the inspection intensity level and execution node defined in the dynamic quality inspection checklist, and combined with the expected output in the production plan, calculate the estimated total execution cost required to implement the corresponding quality inspection strategy.

[0042] For each damage mode family, historical quality loss cost data is collected, the average loss cost of aluminum trim parts defects under that family is calculated, and the total expected quality risk cost is calculated by combining the predicted probability of occurrence of various defects.

[0043] Minimizing the estimated total execution cost and the total expected quality risk cost are the two optimization objectives. The detection nodes, sampling frequency and tolerance standards are the optimization decision variables. A multi-objective optimization problem model is constructed and solved. The solution results are converted into a set of quality inspection performance indicators for the next production cycle.

[0044] The technical effects and advantages of the big data-based quality inspection method for automotive aluminum trim parts of this invention:

[0045] This invention achieves optimized quality management and inspection decisions across the entire process of automotive aluminum trim parts by unifying and correlating vehicle after-sales damage data, production process time-series data, and quality cost information through modeling and correlation analysis. By automatically identifying damage patterns and extracting key process characteristics, this invention can proactively detect potential quality risks in different production batches from a management decision-making perspective, avoiding the lag caused by relying solely on post-production inspections or empirical rules for quality inspection arrangements. Simultaneously, this invention introduces defect propagation modeling and probability prediction mechanisms, transforming quality risk assessment from static judgment to quantifiable and comparable risk metrics, helping management to rationally balance costs and quality. Furthermore, by incorporating quality inspection execution costs and historical quality loss costs into a unified multi-objective optimization framework, this invention can output quality inspection performance indicators for the next production cycle, enabling refined allocation of inspection resources and reducing waste and risks caused by over- or under-inspection.

[0046] Compared with existing methods, this invention is more in line with the actual needs of enterprise quality management and operational decision-making, improves the dynamic adaptability and economy of quality inspection strategies, and provides automotive aluminum trim manufacturing enterprises with a quality decision support solution that has management value and is feasible to implement. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of a quality inspection method for automotive aluminum trim parts based on big data according to the present invention. Detailed Implementation

[0048] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] Figure 1 This invention presents a quality inspection method for automotive aluminum trim parts based on big data, which includes the following steps:

[0051] S1. Extract damage records containing batch identifiers of aluminum trim parts from vehicle after-sales maintenance data sources, perform data cleaning on the damage records, and generate structured batch damage data.

[0052] S2. Perform unsupervised learning based on spectral clustering on batch damage data to classify the damage pattern groups of aluminum decorative parts.

[0053] S3. Based on the batch identifier of the aluminum trim parts, trace the time-series production process data of the corresponding production cycle, and cut it into production segment feature sequences using the sliding window method;

[0054] S4. For each damage pattern family, input the feature sequences of all batches of production fragments contained therein into the integrated classification framework to identify key process feature combinations.

[0055] S5. Based on the combination of key production process characteristics, construct an early propagation network for aluminum decorative parts process defects to predict the types of measurable defects and their probability of occurrence.

[0056] S6. Dynamically generate multi-level hierarchical quality inspection strategies based on the predicted defect types and occurrence probabilities.

[0057] S7. Integrate the estimated execution cost of multi-level hierarchical quality inspection strategies with historical quality loss cost data corresponding to each damage mode group, and output a set of quality inspection performance indicators for the next production cycle through a dynamic bi-objective optimization algorithm.

[0058] In step S1, structured batch damage data is generated.

[0059] Historical data generated during vehicle after-sales maintenance is uniformly accessed and processed. The data source is the actual maintenance business data generated and retained long-term during the vehicle after-sales maintenance process. This data source contains maintenance records generated when vehicles enter the maintenance process due to appearance, assembly, or functional abnormalities during their service life. Each maintenance record corresponds to a specific vehicle, maintenance time, and information on the parts involved. For aluminum trim parts, a specific component type, the records in the after-sales maintenance data source are scanned and filtered one by one, extracting only maintenance records related to actual damage to aluminum trim parts. To ensure the traceability and consistency of subsequent data processing, each damage record is explicitly required to include at least the batch identification information of the aluminum trim part and a textual description of the damage state during the extraction process. The batch identification is used to uniquely identify the manufacturing batch to which the aluminum trim part belongs during the production stage, and the damage morphology description reflects the actual damage manifestation of the aluminum trim part during the after-sales stage. After extraction, all eligible maintenance records are organized and stored in the original business time order, maintaining the original semantic integrity and temporal correlation of the damage records.

[0060] After extracting the damage records, a systematic data cleaning and standardization process is performed on the obtained damage records. This embodiment addresses the issue of format differences in batch identifiers for aluminum trim parts by unifying the data. Since after-sales maintenance data comes from different maintenance outlets and different data entry personnel, batch identifiers may exhibit inconsistencies in capitalization, separators, or additional explanatory characters in the actual records. Therefore, during the cleaning process, the batch identifiers are standardized to ensure that aluminum trim parts from the same production batch have a unique and stable identifier at the data level. Subsequently, the damage morphology descriptions are standardized and mapped. Specifically, the damage phenomena described in natural language in the maintenance records are compared item by item with a pre-organized and fixed damage morphology classification table. Descriptions with consistent or highly similar semantics are uniformly mapped to corresponding standardized damage morphology codes. For example, different expressions such as surface scratches, paint peeling, and deformation are uniformly categorized into clear damage morphology categories. After completing the above cleaning and mapping process, all cleaned damage records are integrated and merged according to the batch identifiers of the aluminum trim parts. Multiple damage records appearing in the same batch are summarized and associated to form a data structure indexed by the part batch. In this structured damage dataset, each batch index is associated with its corresponding set of damage morphology codes and their occurrence, thus forming the basic data format required for subsequent damage pattern analysis and production process correlation analysis.

[0061] In S2, damage mode groups are defined for aluminum trim parts.

[0062] Based on the structured damage dataset, the damage morphology codes corresponding to each batch of aluminum trim parts are statistically analyzed and organized. The damage morphology codes appearing in each batch during the after-sales stage are counted class by class, forming a frequency description result that reflects the damage distribution characteristics of that batch. This frequency description result uses damage morphology category as the dimension and the frequency of occurrence as the value, completely preserving the relative distribution relationship of different damage morphologies within the same batch. On this basis, the damage frequency description results of any two different batches of aluminum trim parts are compared and calculated. Cosine similarity is used to quantify the similarity between the two batches in terms of damage morphology distribution structure. This similarity calculation process uses a multi-dimensional vector composed of damage morphology frequencies as input, and the similarity result is directly derived from the actual damage data itself. After completing the pairwise similarity calculation between batches, the similarity results obtained between all batches of aluminum trim parts are uniformly arranged according to batch order to construct a symmetrical batch similarity matrix. The rows and columns of this similarity matrix correspond to specific batches of aluminum trim parts, and the matrix elements in the same position represent the degree of similarity in damage morphology distribution between the corresponding two batches. A Laplacian matrix is ​​constructed based on the similarity matrix to reflect the overall correlation structure between batches. Eigenvalue decomposition is performed on the Laplacian matrix, and eigenvectors corresponding to several eigenvalues ​​that meet a preset requirement are selected from the decomposition results. These eigenvectors are then combined according to the batch dimension to form a new low-dimensional feature representation space. This new feature space can reduce data dimensionality and weaken noise interference while preserving the original similarity relationship structure, making the positional relationships of different aluminum trim parts batches more clearly distinguishable within this space.

[0063] After obtaining the new feature space, the representations of all aluminum trim parts batches in this feature space are used as input to perform cluster analysis. Unsupervised clustering is employed to group batch features, automatically grouping closely related batches into the same set by calculating the distance relationships between batches in the feature space. The clustering process does not rely on predefined damage category labels but is entirely based on the distribution structure of batches in the feature space, ensuring that the clustering results accurately reflect the inherent patterns of damage morphology distribution. After clustering, several mutually exclusive batch sets are obtained, with each set showing significant consistency in damage morphology distribution among the aluminum trim parts batches. Subsequently, each batch set is defined as a damage pattern family, and a unique family identifier is assigned to each family for distinguishing and referencing different damage patterns in subsequent production process analysis, key feature identification, and quality inspection strategy generation.

[0064] In step S3, based on the batch identifier of the aluminum trim parts, the time-series production process data of the corresponding production cycle is traced, and the sliding window method is used to cut it into a production segment feature sequence.

[0065] Based on the batch identifiers bound to aluminum trim parts during the production stage, a systematic retrieval of production process data recorded in historical production lines is performed. Using the batch identifier as a unique index, multi-dimensional time-series data of that batch along the complete production chain is retrieved from the production execution records. This multi-dimensional time-series data originates from process parameter data continuously collected and recorded during actual production at each manufacturing process, covering continuous operational status information at multiple process nodes. To ensure consistency and splicability of data across different processes in the time dimension, after retrieval, the obtained multi-dimensional time-series data is organized sequentially according to manufacturing process nodes, and the data segments corresponding to each process are aligned according to the predetermined process sequence in the production flow. During alignment, the actual entry and exit times of processes are used as a reference to completely preserve data belonging to the same process node, while eliminating time misalignment issues caused by differences in data acquisition frequency or recording delays. After completing process-level alignment, the time-series data corresponding to each process node are spliced ​​according to the actual sequence of the production chain, forming a continuous data stream covering the batch from the initial process to the final process. This continuous data stream maintains a monotonically increasing trend in the time dimension and fully reflects the entire batch production process in the process dimension, thus constituting a global production process data stream that can characterize the overall production behavior of the batch.

[0066] After obtaining the global production process data stream corresponding to a batch, the data stream is segmented based on the time axis to form a production segment feature sequence. Using the theoretical production cycle time corresponding to the completion of all manufacturing processes for a single aluminum trim product on the production line as a reference benchmark, the time width and step interval of the sliding window are defined. The window width and step interval are determined based on the production cycle time and process continuity. For example, the window width is set to cover the time length of several consecutive process runs, and the step interval is set to a certain proportion of the window width to ensure continuity between adjacent windows rather than complete separation. The data stream is continuously segmented along the time axis of the global production process data stream, with each segment generating a production data segment within a corresponding time period. For each production data segment, statistical feature calculations are performed on the various production process parameters contained within it. These statistical features include descriptive indicators that reflect the overall level and fluctuation state of the parameters, thereby converting the original high-frequency time series data into a structured feature representation. After completing the feature calculation for a single window, the statistical feature results corresponding to all sliding windows within the same batch are arranged and combined in chronological order to form a production segment feature sequence consistent with the time evolution order. This production segment feature sequence, while maintaining temporal continuity, enables segmented characterization of the dynamic changes in the production process.

[0067] In step S4, for each damage pattern family, the feature sequences of all batches of production fragments contained therein are input into the integrated classification framework to identify key process feature combinations.

[0068] Based on the completed damage pattern family classification, historical production data is further organized and filtered to form a sample set for classification modeling. Damage pattern families are used as the basic unit for sample classification, and each family is processed separately. According to the batch identifier of the aluminum trim parts, all batches belonging to that damage pattern family are retrieved one by one, and the production segment feature sequences formed during the production stage of these batches are extracted. These production segment feature sequences are generated using a sliding window method, fully reflecting the process feature changes of the corresponding batch at each time segment within the production cycle. Based on this, the production segment feature sequences corresponding to all batches within a family are uniformly grouped into a positive sample set. Each sample in the positive sample set clearly corresponds to a specific batch and its time-sequential production segment features. Subsequently, aluminum trim part batches not classified into any damage pattern family are selected from the historical production data. These batches have no clear damage records during the after-sales stage and are not associated with any damage morphology codes in the structured damage dataset. The production segment feature sequences corresponding to these batches are extracted and integrated as a negative sample set. By using the above methods, we can ensure that the positive and negative sample sets are consistent in terms of data source, feature composition, and temporal representation, with differences only in the attribution of damage patterns. This provides a clear and distinguishable sample basis for the construction of subsequent classification models.

[0069] After constructing the positive and negative sample sets, the two types of samples are organized uniformly to form a dataset for model training. The ensemble classification model consists of multiple decision structure-based classification units. Each classification unit learns to divide sample features during training, and the results are fused at the model level. During model training, production segment feature sequences are used as input data. Each input sample contains multiple production segment features arranged in chronological order, used to characterize the process state evolution of the corresponding part batch throughout the entire production cycle. The model output is a batch classification prediction result, used to determine the damage pattern family to which the input sample belongs or whether it belongs to that family. For example, for a certain damage pattern family, 120 aluminum trim part batches within that family are selected as positive samples, while 180 batches from normal batches that have never been classified into any damage pattern family are selected as negative samples. The production segment feature sequence corresponding to each batch consists of 10 consecutive sliding windows. Each window contains statistical features calculated from production parameters such as temperature, pressure, and speed within the window. During model training, an ensemble classification model consisting of a random forest classifier and a gradient boosting decision tree classifier was used. The random forest component contained 100 decision trees, with a maximum depth of 8 for each tree. The gradient boosting decision tree component was set to 80 iterations, with a single-element learning step size of 0.1. During training, positive and negative samples were divided into training and validation data in an 8:2 ratio. The model iteratively trained using batch classification results as the output target. Training was terminated when the classification accuracy on the validation set no longer improved after 5 consecutive iterations, resulting in the final ensemble classification model used for batch classification prediction.

[0070] During training, samples from both the positive and negative sample sets are simultaneously input into the model. Each classification unit continuously adjusts its internal decision structure based on the differences in sample features during the training phase, enabling the model to gradually develop the ability to distinguish the characteristics of different damage patterns. The training process strictly maintains a consistent mapping relationship between input features and output classifications, ensuring that after training, the model can provide stable and reproducible batch classification prediction results for any input production fragment feature sequence.

[0071] After training the ensemble classification model, the decision structure formed within the model is analyzed to identify the process features that play a major role in the classification results. For all process features used during model training, their usage and participation in sample splitting at each classification decision node are statistically analyzed. Specifically, the contribution of each process feature to the sample discrimination results during decision node splitting in the model training phase is cumulatively recorded. This contribution reflects the strength of the actual discriminative role of the feature in distinguishing positive and negative samples. This is achieved by recording the number of times each process feature is used in split nodes in the random forest and gradient boosting decision tree, and simultaneously accumulating the impact of each split on the class purity of the positive and negative sample discrimination results. The cumulative contributions of the same process feature across all decision trees and iterations are summarized to form a comprehensive contribution value for the feature, and the results are ranked by importance. Class purity refers to the concentration of positive and negative sample distributions in the sample set. Its influence is measured by comparing the reduction in uncertainty of the class distribution in the sample set before and after splitting; the greater the reduction in uncertainty after splitting, the greater the contribution of the process feature to sample discrimination. Subsequently, based on this importance ranking, a subset of features with the highest contribution levels is selected from all process features. This selection process prioritizes coverage of key discriminative information while avoiding the introduction of redundant or noisy features. Finally, the selected feature subset is defined as the key process feature combination corresponding to the damage mode family. This key process feature combination is structurally directly related to specific production processes and parameters, and semantically reflects the abnormal production process features highly correlated with the damage mode.

[0072] In step S5, an early propagation network for process defects in aluminum decorative parts is constructed to predict the types of measurable defects and their probability of occurrence.

[0073] The identified key process feature combinations are mapped at the process level. Each feature in the key process feature combination originates from the statistical performance of specific production parameters under a specific manufacturing process. Therefore, during implementation, based on the correspondence between parameters and processes in the production process records, the parameters involved in the calculation of each key process feature are explicitly associated with the manufacturing process node that generates it. After mapping, each manufacturing process involved in the complete production chain of aluminum trim parts is defined as a node in a graph network, with each node carrying key process feature information associated with that process. After the nodes are constructed, directed connections are established between nodes one by one according to the sequential relationship of processes in the actual production flow of aluminum trim parts. These directed connections are used to represent the possible propagation direction of defects in the process path. The directed graph network model constructed in this way structurally reflects the processing sequence of aluminum trim parts from the front-end process to the back-end process, and semantically clearly describes the path constraints of defects gradually propagating along the process flow.

[0074] After constructing the directed graph network model, historical production data and inspection data are jointly analyzed to quantify the transfer relationship of defect states between adjacent manufacturing nodes. Specifically, for any pair of directly connected adjacent manufacturing nodes in the directed graph network, the occurrence of a specific defect detected in the downstream node is statistically analyzed in historical production batches when the key process characteristics corresponding to the upstream node exhibit abnormal behavior. This statistical process is conducted on a batch-by-batch basis, simultaneously checking whether each batch exhibits the corresponding key process characteristics in the upstream process stage and whether a clear defect type is recorded in subsequent inspection stages. By accumulating the number of historical batches meeting these conditions and comparing it with the total number of batches exhibiting the upstream key process characteristics, the conditional probability value of a specific defect state occurring in the downstream node under given upstream key process characteristic states is calculated. This conditional probability value is recorded as an attribute of the corresponding edge in the directed graph network model, used to describe the strength and probability of defect propagation between adjacent process nodes.

[0075] When a batch of aluminum trim parts enters the analysis stage, the feature sequence of the corresponding production segment is extracted, and feature state information matching the key process feature combination is identified. Subsequently, this production segment feature sequence is loaded as input into a pre-constructed directed graph network model, and forward propagation reasoning is performed sequentially from upstream process nodes to downstream process nodes according to the actual manufacturing process sequence. In specific implementation, at each manufacturing node, the probability of a defect occurring is progressively updated based on the key process feature state carried by that node and the conditional probability information recorded on its connected edges, and the updated defect probability is passed to subsequent nodes. Through continuous propagation and cumulative calculation along the process path, the probability of various defects being detected in each downstream manufacturing process and corresponding inspection stage is finally obtained. After completing the full-path reasoning, the defect occurrence probability is compared with a pre-set confidence standard, where the confidence standard is set based on historical inspection stability data; for example, a confidence level of 0.7 is used to screen defect types with a high probability of occurrence. Finally, the defect types with a probability exceeding the confidence level and their corresponding probability values ​​are used as the defect prediction output for this batch, providing probability input for the generation of quality inspection strategies.

[0076] In step S6, a multi-level hierarchical quality inspection strategy is dynamically generated based on the predicted defect type and its probability of occurrence.

[0077] Based on the obtained defect prediction results, the occurrence probability of each defect type is uniformly processed and classified according to a pre-defined risk level range. The risk level range is set based on the company's historical quality data and quality inspection management experience. For example, the defect occurrence probability is divided into low-risk, medium-risk, and high-risk ranges from low to high, and the boundaries of each range are clearly defined and stabilized before implementation. After completing the risk level classification, a corresponding detection intensity level is matched for each risk level range. The detection intensity level is defined by both the detection type and the sampling benchmark ratio, and the specific quality inspection step for implementing this detection intensity is also specified. The detection type indicates the rigor of the detection method, and the sampling benchmark ratio limits the proportion of products to be sampled in the corresponding quality inspection step. When the predicted occurrence probability of a surface scratch defect is classified into the low-risk range, the corresponding matched detection intensity level is routine visual inspection, and the sampling benchmark ratio is set at a low level, such as sampling at a fixed ratio in the final visual inspection process. When the predicted probability of a structural deformation defect is classified as high-risk, the corresponding matching inspection intensity level is enhanced inspection. The inspection type includes a combination of dimensional measurement and visual inspection, the sampling benchmark ratio is set at a high level, and inspection is explicitly performed simultaneously in the finished product inspection process and the preceding key processes. Through the above methods, a clear distinction is made in the allocation of inspection resources for defects of different risk levels.

[0078] After matching the risk level of each defect type with the inspection intensity level, all predicted defect types involved in the current batch of aluminum trim parts are comprehensively compiled. The risk level information corresponding to each defect type in this batch is read one by one, and the inspection intensity level and the specific quality inspection stage associated with that risk level are retrieved simultaneously. Subsequently, the above information is summarized according to the defect type dimension, presenting multiple defect risks that may be involved in the same batch side-by-side, forming a quality inspection decision result covering multiple risk levels. During the integration process, situations where multiple inspection requirements exist under the same quality inspection stage are uniformly merged to ensure that implementation is carried out according to the highest corresponding inspection intensity level at the execution level. Finally, the integrated defect type, risk level, inspection intensity level, and quality inspection stage information are structured in the form of a checklist, generating a dynamic quality inspection checklist with multi-level hierarchical characteristics. The dynamic quality inspection checklist clearly lists the types of tests to be performed at each quality inspection stage and the corresponding sampling benchmark ratio. In terms of sequence, it reflects the hierarchical relationship from high risk to low risk, thereby forming a clear and executable hierarchical quality inspection arrangement within the same production batch, ensuring the consistency and operability of the quality inspection execution process.

[0079] In step S7, the set of quality inspection performance indicators for the next production cycle is output.

[0080] Based on the generated dynamic quality inspection checklist, the cost of all quality inspection activities required in the current production cycle is calculated. Specifically, the inspection intensity level information for each quality inspection step is retrieved from the dynamic quality inspection checklist, including the specific inspection type and the corresponding sampling benchmark ratio. Simultaneously, combined with the projected output data already specified in the production plan, the number of products that each quality inspection step needs to cover in this production cycle is confirmed. Then, for each quality inspection step, according to the actual inspection method corresponding to its inspection type, the unit execution cost record of that inspection method in historical execution is retrieved. This unit execution cost includes actual cost elements such as manual inspection hours, equipment occupancy time, and necessary consumable consumption. After determining the unit execution cost, based on the sampling benchmark ratio and projected output of that quality inspection step, the number of inspections required for that step in the entire production cycle is calculated, and the total execution cost corresponding to that quality inspection step is obtained. After completing the above cost calculation for each quality inspection step listed in the dynamic quality inspection checklist, the execution costs of each quality inspection step are summed to obtain the estimated total execution cost required to implement the current dynamic quality inspection strategy.

[0081] A systematic quantification of quality risk costs is performed. For each damage mode family, quality loss records related to that family are extracted from the company's historical quality management data. These records include actual economic losses incurred due to defect rework, scrapping, claims, etc. By organizing and statistically analyzing historical data, the average quality loss cost corresponding to defects in a single aluminum trim part under that damage mode family is calculated. After calculating the average loss cost, the probability information of various defects obtained in the aforementioned defect prediction step is used to estimate the potential quality risks in the current production cycle. The probability of occurrence of various defects under the same damage mode family is correlated with their corresponding average quality loss cost to calculate the expected quality risk cost that family may bring under the current production conditions. After performing the above calculations for all damage mode families, the expected quality risk costs corresponding to each family are summarized to obtain the total expected quality risk cost under the current dynamic quality inspection strategy.

[0082] After obtaining the estimated total execution cost and the total expected quality risk cost, the quality inspection strategy is optimized. In this embodiment, minimizing the estimated total execution cost and minimizing the total expected quality risk cost are two simultaneous and mutually constraining optimization objectives, constructing a multi-objective optimization problem model. In this optimization model, the inspection nodes involved in the dynamic quality inspection checklist, the sampling frequency corresponding to each inspection node, and the tolerance standards used in the inspection process are used as optimization decision variables. In specific implementation, while ensuring that the inspection intensity level and process constraints are not violated, the value range of the above decision variables is limited, and the parameter combination that achieves a reasonable balance between cost input and quality risk is searched through the multi-objective optimization solution process. After the solution is completed, a solution that meets the current production management objectives is selected from the obtained Pareto optimal solution set, and the corresponding inspection node configuration, sampling frequency adjustment results, and tolerance standard adjustment are organized and output. Finally, the above parameter combination is converted into the quality inspection performance indicators for the next production cycle in a structured form, which are used to control the production standards for aluminum trim parts quality.

[0083] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0084] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0087] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0089] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0091] 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 quality inspection method for automotive aluminum trim parts based on big data, characterized in that, Includes the following steps: S1. Extract damage records containing batch identifiers of aluminum trim parts from vehicle after-sales maintenance data sources, perform data cleaning on the damage records, and generate structured batch damage data. S2. Perform unsupervised learning based on spectral clustering on batch damage data to classify the damage pattern groups of aluminum decorative parts. S3. Based on the batch identifier of the aluminum trim parts, trace the time-series production process data of the corresponding production cycle, and cut it into production segment feature sequences using the sliding window method; S4. For each damage pattern family, input the feature sequences of all batches of production fragments contained therein into the integrated classification framework to identify key process feature combinations. S5. Based on the combination of key production process characteristics, construct an early propagation network for aluminum decorative parts process defects to predict the types of measurable defects and their probability of occurrence. S6. Dynamically generate multi-level hierarchical quality inspection strategies based on the predicted defect types and occurrence probabilities. S7. Integrate the estimated execution cost of multi-level hierarchical quality inspection strategies with historical quality loss cost data corresponding to each damage mode group, and output a set of quality inspection performance indicators for the next production cycle through a dynamic bi-objective optimization algorithm. In step S4, for each damage pattern family, the feature sequences of all batches of production fragments contained therein are input into the integrated classification framework to identify key process feature combinations, specifically including: For each damage mode family, the production segment feature sequence corresponding to the batch identifier of all aluminum trim parts in the family is used as the positive sample set. Normal parts batches that are not classified into any damage mode family are extracted from historical production data, and the corresponding production segment feature sequence is used as the negative sample set. An ensemble classification model is trained using positive and negative sample sets, taking production fragment feature sequences as input and batch classification prediction as output. The cumulative contribution of each process feature to the sample discrimination result when participating in the classification decision node division during model training is statistically analyzed. The importance ranking of the process features on which the integrated classification model depends in the classification decision is calculated, and the subset of features with the highest importance is selected as the key process feature combination. In step S5, constructing an early propagation network for process defects in aluminum decorative parts and predicting the types and probabilities of measurable defects specifically includes: Key process features are combined and mapped to corresponding manufacturing nodes. A directed graph network model is constructed based on the sequential relationship of manufacturing nodes. Nodes represent manufacturing operations that carry key process features, and edges represent the propagation relationship of defects along the process path. By utilizing the co-occurrence relationship between the same key process features extracted from historical production data and the final detected defects, the conditional probability of defect state transition between adjacent nodes in the directed graph network model is calculated. In the current batch of aluminum decorative parts, the feature sequence of production segments containing key process characteristics is used as input. Forward propagation reasoning is performed along the process path in the directed graph network model to calculate the probability of various defects that are finally detected in each downstream process in real time. The defect types with probabilities exceeding the set confidence level and their corresponding probability values ​​are used as prediction outputs.

2. The quality inspection method for automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S1, generating structured batch damage data specifically includes: Access vehicle after-sales maintenance data sources and extract damage records for aluminum trim parts. The records include batch identifiers of the aluminum trim parts and descriptions of the damage morphology. Data cleaning is performed on the extracted damage records to unify the format of part batch identifiers and map the damage morphology descriptions to standardized damage morphology codes. The damage records after cleaning are integrated according to the batch identifier of the aluminum trim parts to generate a structured damage dataset indexed by the part batch.

3. The quality inspection method for automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S2, the specific classification of damage mode groups for aluminum trim parts includes: Based on the frequency of occurrence of various damage morphology codes of each batch of aluminum trim parts in the structured damage dataset, calculate the cosine similarity between any two batches of aluminum trim parts. Construct a symmetric batch similarity matrix for parts and calculate the corresponding Laplacian matrix. Perform eigenvalue decomposition on the Laplacian matrix and select the eigenvectors corresponding to the first few largest eigenvalues ​​to form a new feature space. In the new feature space, a clustering algorithm is used to divide all aluminum trim parts into several mutually exclusive sets. Each set is defined as a damage pattern family and labeled with a family identifier.

4. The quality inspection method for automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S3, based on the batch identifier of the aluminum trim parts, the time-series production process data corresponding to the production cycle is traced, and the sliding window method is used to cut it into a production segment feature sequence, specifically including: Based on the batch identifier of aluminum trim parts, retrieve multi-dimensional time series data of each batch recorded along the complete production chain from historical production lines; The retrieved multidimensional time series data are aligned and stitched together according to the manufacturing process nodes to form a continuous global production process data stream for this batch; Based on the theoretical production cycle time of a single product, the width and step interval of the sliding window are defined, and the global production process data stream is continuously captured along the time axis. For each data segment captured by the sliding window, calculate the statistical characteristics of various production parameters contained therein within the time window. Arrange the statistical characteristics of all windows in the batch in chronological order to form the production segment feature sequence corresponding to this batch.

5. The quality inspection method for automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S6, the dynamic generation of a multi-level hierarchical quality inspection strategy based on the predicted defect type and probability of occurrence specifically includes: The probability of occurrence for each predicted defect type is divided into different risk level ranges; For each risk level range, a preset detection intensity level is matched. The detection intensity level defines the detection type and sampling benchmark ratio corresponding to the risk level defect, and at the same time determines the quality inspection process to be performed for the corresponding detection intensity level. By integrating the risk level, inspection intensity level, and quality inspection information corresponding to all predicted defect types in the current batch of aluminum trim parts, a dynamic quality inspection checklist with multi-level hierarchical quality inspection strategies is generated.

6. The quality inspection method for automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S7, the set of quality inspection performance indicators for the next production cycle is output, specifically including: Based on the inspection intensity level and execution node defined in the dynamic quality inspection checklist, and combined with the expected output in the production plan, calculate the estimated total execution cost required to implement the corresponding quality inspection strategy. For each damage mode family, historical quality loss cost data is collected, the average loss cost of aluminum trim parts defects under that family is calculated, and the total expected quality risk cost is calculated by combining the predicted probability of occurrence of various defects. Minimizing the estimated total execution cost and the total expected quality risk cost are the two optimization objectives. The detection nodes, sampling frequency and tolerance standards are the optimization decision variables. A multi-objective optimization problem model is constructed and solved. The solution results are converted into a set of quality inspection performance indicators for the next production cycle.

Images